x/imagegen: limit input images to 2 and remove obvious comment

- Add OpenAI-compatible /v1/images/edits endpoint
- Pass input images through the generation pipeline
- Add ImageEditModel interface for models that support image editing
- Return error when images provided but model doesn't support editing

api/types.go

@@ -749,7 +749,7 @@ type ShowResponse struct {
 	Messages      []Message          `json:"messages,omitempty"`
 	RemoteModel   string             `json:"remote_model,omitempty"`
 	RemoteHost    string             `json:"remote_host,omitempty"`
-	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`

cmd/cmd.go

@@ -101,6 +101,67 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return fmt.Errorf("invalid model name: %s", modelName)
 	}
 
+	// Check for --experimental flag for safetensors model creation
+	experimental, _ := cmd.Flags().GetBool("experimental")
+	if experimental {
+		// Get Modelfile content - either from -f flag or default to "FROM ."
+		var reader io.Reader
+		filename, err := getModelfileName(cmd)
+		if os.IsNotExist(err) || filename == "" {
+			// No Modelfile specified or found - use default
+			reader = strings.NewReader("FROM .\n")
+		} else if err != nil {
+			return err
+		} else {
+			f, err := os.Open(filename)
+			if err != nil {
+				return err
+			}
+			defer f.Close()
+			reader = f
+		}
+
+		// Parse the Modelfile
+		modelfile, err := parser.ParseFile(reader)
+		if err != nil {
+			return fmt.Errorf("failed to parse Modelfile: %w", err)
+		}
+
+		// Extract FROM path and configuration
+		var modelDir string
+		mfConfig := &xcreateclient.ModelfileConfig{}
+
+		for _, cmd := range modelfile.Commands {
+			switch cmd.Name {
+			case "model":
+				modelDir = cmd.Args
+			case "template":
+				mfConfig.Template = cmd.Args
+			case "system":
+				mfConfig.System = cmd.Args
+			case "license":
+				mfConfig.License = cmd.Args
+			}
+		}
+
+		if modelDir == "" {
+			modelDir = "."
+		}
+
+		// Resolve relative paths based on Modelfile location
+		if !filepath.IsAbs(modelDir) && filename != "" {
+			modelDir = filepath.Join(filepath.Dir(filename), modelDir)
+		}
+
+		quantize, _ := cmd.Flags().GetString("quantize")
+		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+			ModelName: modelName,
+			ModelDir:  modelDir,
+			Quantize:  quantize,
+			Modelfile: mfConfig,
+		}, p)
+	}
+
 	var reader io.Reader
 
 	filename, err := getModelfileName(cmd)
@@ -136,28 +197,6 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	// Check if this is a tensor model (image generation) and handle it directly
-	quantize, _ := cmd.Flags().GetString("quantize")
-	modelDir := filepath.Dir(filename)
-	for _, cmd := range modelfile.Commands {
-		if cmd.Name == "model" {
-			if filepath.IsAbs(cmd.Args) {
-				modelDir = cmd.Args
-			} else {
-				modelDir = filepath.Join(filepath.Dir(filename), cmd.Args)
-			}
-			break
-		}
-	}
-	if create.IsTensorModelDir(modelDir) {
-		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
-			ModelName: modelName,
-			ModelDir:  modelDir,
-			Quantize:  quantize,
-			Modelfile: xcreateclient.ExtractModelfileConfig(modelfile),
-		}, p)
-	}
-
 	status := "gathering model components"
 	spinner := progress.NewSpinner(status)
 	p.Add(status, spinner)
@@ -169,6 +208,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	spinner.Stop()
 
 	req.Model = modelName
+	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
 	}
@@ -1775,15 +1815,22 @@ func NewCLI() *cobra.Command {
 	rootCmd.Flags().BoolP("version", "v", false, "Show version information")
 
 	createCmd := &cobra.Command{
-		Use:     "create MODEL",
-		Short:   "Create a model",
-		Args:    cobra.ExactArgs(1),
-		PreRunE: checkServerHeartbeat,
-		RunE:    CreateHandler,
+		Use:   "create MODEL",
+		Short: "Create a model",
+		Args:  cobra.ExactArgs(1),
+		PreRunE: func(cmd *cobra.Command, args []string) error {
+			// Skip server check for experimental mode (writes directly to disk)
+			if experimental, _ := cmd.Flags().GetBool("experimental"); experimental {
+				return nil
+			}
+			return checkServerHeartbeat(cmd, args)
+		},
+		RunE: CreateHandler,
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
+	createCmd.Flags().Bool("experimental", false, "Enable experimental safetensors model creation")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",

convert/convert.go

@@ -313,6 +313,8 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &deepseek2Model{}
 	case "Glm4MoeLiteForCausalLM":
 		conv = &glm4MoeLiteModel{}
+	case "Lfm2ForCausalLM":
+		conv = &lfm2Model{}
 	default:
 		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_lfm2.go

@@ -0,0 +1,100 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type lfm2Model struct {
+	ModelParameters
+	HiddenSize            uint32   `json:"hidden_size"`
+	NumHiddenLayers       uint32   `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32   `json:"max_position_embeddings"`
+	IntermediateSize      uint32   `json:"intermediate_size"`
+	NumAttentionHeads     uint32   `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32   `json:"num_key_value_heads"`
+	RopeTheta             float32  `json:"rope_theta"`
+	NormEps               float32  `json:"norm_eps"`
+	ConvLCache            uint32   `json:"conv_L_cache"`
+	LayerTypes            []string `json:"layer_types"`
+	TieEmbedding          bool     `json:"tie_embedding"`
+}
+
+var _ ModelConverter = (*lfm2Model)(nil)
+
+func (p *lfm2Model) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "lfm2"
+	kv["lfm2.vocab_size"] = p.VocabSize
+	kv["lfm2.block_count"] = p.NumHiddenLayers
+	kv["lfm2.embedding_length"] = p.HiddenSize
+	kv["lfm2.feed_forward_length"] = p.IntermediateSize
+	kv["lfm2.context_length"] = p.MaxPositionEmbeddings
+
+	// Build per-layer KV head count array based on layer_types
+	// (0 = shortconv layer, non-zero = attention layer with that many KV heads)
+	kvHeadCounts := make([]uint32, p.NumHiddenLayers)
+	for i := range p.NumHiddenLayers {
+		if int(i) < len(p.LayerTypes) && p.LayerTypes[i] == "full_attention" {
+			kvHeadCounts[i] = p.NumKeyValueHeads
+		}
+	}
+
+	kv["lfm2.attention.head_count"] = p.NumAttentionHeads
+	kv["lfm2.attention.head_count_kv"] = kvHeadCounts
+	kv["lfm2.attention.key_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.value_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.layer_norm_rms_epsilon"] = p.NormEps
+	kv["lfm2.rope.freq_base"] = p.RopeTheta
+	kv["lfm2.shortconv.l_cache"] = p.ConvLCache
+
+	return kv
+}
+
+func (p *lfm2Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		shape := t.Shape()
+
+		// Squeeze conv weights: [D, 1, K] -> [D, K]
+		if strings.HasSuffix(t.Name(), "shortconv.conv.weight") {
+			if len(shape) == 3 && shape[1] == 1 {
+				shape = []uint64{shape[0], shape[2]}
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    slices.Clone(shape),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *lfm2Model) Replacements() []string {
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.embedding_norm", "output_norm",
+		"model.layers", "blk",
+		"operator_norm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.out_proj", "attn_output",
+		"self_attn.q_layernorm", "attn_q_norm",
+		"self_attn.k_layernorm", "attn_k_norm",
+		"conv.conv", "shortconv.conv",
+		"conv.in_proj", "shortconv.in_proj",
+		"conv.out_proj", "shortconv.out_proj",
+		"feed_forward.w1", "ffn_gate",
+		"feed_forward.w2", "ffn_down",
+		"feed_forward.w3", "ffn_up",
+		"ffn_norm", "ffn_norm",
+	}
+}

convert/reader.go

@@ -40,6 +40,7 @@ const (
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
 		strings.HasSuffix(t.name, ".bias") ||
+		strings.HasSuffix(t.name, ".shortconv.conv.weight") ||
 		t.name == "token_types.weight" ||
 		t.name == "v.positional_embedding_vlm" ||
 		t.name == "v.tile_position_embd.weight" ||

fs/ggml/ggml.go

@@ -270,6 +270,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 		"glm4moelite",
+		"lfm2",
 	}, kv.Architecture())
 }
 
@@ -859,6 +860,7 @@ func (f GGML) FlashAttention() bool {
 		"gemma3",
 		"glm4moelite",
 		"gptoss", "gpt-oss",
+		"lfm2",
 		"mistral3",
 		"olmo3",
 		"qwen3", "qwen3moe",

integration/imagegen_test.go

@@ -0,0 +1,148 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"encoding/base64"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestImageGeneration(t *testing.T) {
+	skipUnderMinVRAM(t, 8)
+
+	type testCase struct {
+		imageGenModel string
+		visionModel   string
+		prompt        string
+		expectedWords []string
+	}
+
+	testCases := []testCase{
+		{
+			imageGenModel: "jmorgan/z-image-turbo",
+			visionModel:   "llama3.2-vision",
+			prompt:        "A cartoon style llama flying like a superhero through the air with clouds in the background",
+			expectedWords: []string{"llama", "flying", "cartoon", "cloud", "sky", "superhero", "air", "animal", "camelid"},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(fmt.Sprintf("%s->%s", tc.imageGenModel, tc.visionModel), func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+			defer cancel()
+
+			client, _, cleanup := InitServerConnection(ctx, t)
+			defer cleanup()
+
+			// Pull both models
+			if err := PullIfMissing(ctx, client, tc.imageGenModel); err != nil {
+				t.Fatalf("failed to pull image gen model: %v", err)
+			}
+			if err := PullIfMissing(ctx, client, tc.visionModel); err != nil {
+				t.Fatalf("failed to pull vision model: %v", err)
+			}
+
+			// Generate the image
+			t.Logf("Generating image with prompt: %s", tc.prompt)
+			imageBase64, err := generateImage(ctx, client, tc.imageGenModel, tc.prompt)
+			if err != nil {
+				if strings.Contains(err.Error(), "image generation not available") {
+					t.Skip("Target system does not support image generation")
+				} else if strings.Contains(err.Error(), "executable file not found in") { // Windows pattern, not yet supported
+					t.Skip("Windows does not support image generation yet")
+				} else if strings.Contains(err.Error(), "CUDA driver version is insufficient") {
+					t.Skip("Driver is too old")
+				} else if strings.Contains(err.Error(), "insufficient memory for image generation") {
+					t.Skip("insufficient memory for image generation")
+				} else if strings.Contains(err.Error(), "error while loading shared libraries: libcuda.so.1") { // AMD GPU or CPU
+					t.Skip("CUDA GPU is not available")
+				} else if strings.Contains(err.Error(), "ollama-mlx: no such file or directory") {
+					// most likely linux arm - not supported yet
+					t.Skip("unsupported architecture")
+				}
+				t.Fatalf("failed to generate image: %v", err)
+			}
+
+			imageData, err := base64.StdEncoding.DecodeString(imageBase64)
+			if err != nil {
+				t.Fatalf("failed to decode image: %v", err)
+			}
+			t.Logf("Generated image: %d bytes", len(imageData))
+
+			// Preload vision model and check GPU loading
+			err = client.Generate(ctx, &api.GenerateRequest{Model: tc.visionModel}, func(response api.GenerateResponse) error { return nil })
+			if err != nil {
+				t.Fatalf("failed to load vision model: %v", err)
+			}
+
+			// Use vision model to describe the image
+			chatReq := api.ChatRequest{
+				Model: tc.visionModel,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Describe this image in detail. What is shown? What style is it? What is the main subject doing?",
+						Images:  []api.ImageData{imageData},
+					},
+				},
+				Stream: &stream,
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}
+
+			// Verify the vision model's response contains expected keywords
+			response := DoChat(ctx, t, client, chatReq, tc.expectedWords, 240*time.Second, 30*time.Second)
+			if response != nil {
+				t.Logf("Vision model response: %s", response.Content)
+
+				// Additional detailed check for keywords
+				content := strings.ToLower(response.Content)
+				foundWords := []string{}
+				missingWords := []string{}
+				for _, word := range tc.expectedWords {
+					if strings.Contains(content, word) {
+						foundWords = append(foundWords, word)
+					} else {
+						missingWords = append(missingWords, word)
+					}
+				}
+				t.Logf("Found keywords: %v", foundWords)
+				if len(missingWords) > 0 {
+					t.Logf("Missing keywords (at least one was found so test passed): %v", missingWords)
+				}
+			}
+		})
+	}
+}
+
+// generateImage calls the Ollama API to generate an image and returns the base64 image data
+func generateImage(ctx context.Context, client *api.Client, model, prompt string) (string, error) {
+	var imageBase64 string
+
+	err := client.Generate(ctx, &api.GenerateRequest{
+		Model:  model,
+		Prompt: prompt,
+	}, func(resp api.GenerateResponse) error {
+		if resp.Image != "" {
+			imageBase64 = resp.Image
+		}
+		return nil
+	})
+	if err != nil {
+		return "", fmt.Errorf("failed to generate image: %w", err)
+	}
+
+	if imageBase64 == "" {
+		return "", fmt.Errorf("no image data in response")
+	}
+
+	return imageBase64, nil
+}

integration/utils_test.go

@@ -38,6 +38,7 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"lfm2.5-thinking",
 		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
@@ -143,6 +144,7 @@ var (
 		"granite3.3",
 		"hermes3",
 		"internlm2",
+		"lfm2.5-thinking",
 		"llama-guard3",
 		"llama-pro",
 		"llama2-chinese",
@@ -263,6 +265,7 @@ var (
 		"snowflake-arctic-embed2",
 	}
 	libraryToolsModels = []string{
+		"lfm2.5-thinking",
 		"qwen3-vl",
 		"gpt-oss:20b",
 		"gpt-oss:120b",

manifest/layer.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
@@ -14,7 +14,7 @@ type Layer struct {
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
 	Name      string `json:"name,omitempty"` // tensor name, e.g., "text_encoder/model.embed_tokens.weight"
-	status    string
+	Status    string `json:"-"`
 }
 
 const (
@@ -22,7 +22,7 @@ const (
 )
 
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
-	blobs, err := GetBlobsPath("")
+	blobs, err := BlobsPath("")
 	if err != nil {
 		return Layer{}, err
 	}
@@ -45,7 +45,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	}
 
 	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -65,7 +65,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 		MediaType: mediatype,
 		Digest:    digest,
 		Size:      n,
-		status:    fmt.Sprintf("%s %s", status, digest),
+		Status:    fmt.Sprintf("%s %s", status, digest),
 	}, nil
 }
 
@@ -74,7 +74,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		return Layer{}, errors.New("creating new layer from layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -89,7 +89,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		Digest:    digest,
 		Size:      fi.Size(),
 		From:      from,
-		status:    fmt.Sprintf("using existing layer %s", digest),
+		Status:    fmt.Sprintf("using existing layer %s", digest),
 	}, nil
 }
 
@@ -98,7 +98,7 @@ func (l *Layer) Open() (io.ReadSeekCloser, error) {
 		return nil, errors.New("opening layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -126,7 +126,7 @@ func (l *Layer) Remove() error {
 		}
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return err
 	}

manifest/manifest.go

@@ -1,10 +1,9 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -33,12 +32,38 @@ func (m *Manifest) Size() (size int64) {
 	return
 }
 
+func (m *Manifest) Digest() string {
+	return m.digest
+}
+
+func (m *Manifest) FileInfo() os.FileInfo {
+	return m.fi
+}
+
+// ReadConfigJSON reads and unmarshals a config layer as JSON.
+func (m *Manifest) ReadConfigJSON(configPath string, v any) error {
+	for _, layer := range m.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.json" && layer.Name == configPath {
+			blobPath, err := BlobsPath(layer.Digest)
+			if err != nil {
+				return err
+			}
+			data, err := os.ReadFile(blobPath)
+			if err != nil {
+				return err
+			}
+			return json.Unmarshal(data, v)
+		}
+	}
+	return fmt.Errorf("config %q not found in manifest", configPath)
+}
+
 func (m *Manifest) Remove() error {
 	if err := os.Remove(m.filepath); err != nil {
 		return err
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -70,11 +95,11 @@ func (m *Manifest) RemoveLayers() error {
 		if _, used := inUse[layer.Digest]; used {
 			continue
 		}
-		blob, err := GetBlobsPath(layer.Digest)
+		blob, err := BlobsPath(layer.Digest)
 		if err != nil {
 			return err
 		}
-		if err := os.Remove(blob); errors.Is(err, os.ErrNotExist) {
+		if err := os.Remove(blob); os.IsNotExist(err) {
 			slog.Debug("layer does not exist", "digest", layer.Digest)
 		} else if err != nil {
 			return err
@@ -89,7 +114,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, model.Unqualified(n)
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}
@@ -121,7 +146,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 }
 
 func WriteManifest(name model.Name, config Layer, layers []Layer) error {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -148,7 +173,7 @@ func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 }
 
 func Manifests(continueOnError bool) (map[model.Name]*Manifest, error) {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}

manifest/manifest_test.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"encoding/json"

manifest/paths.go

@@ -0,0 +1,95 @@
+package manifest
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
+)
+
+var ErrInvalidDigestFormat = errors.New("invalid digest format")
+
+func Path() (string, error) {
+	path := filepath.Join(envconfig.Models(), "manifests")
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PathForName returns the path to the manifest file for a specific model name.
+func PathForName(n model.Name) (string, error) {
+	if !n.IsValid() {
+		return "", os.ErrNotExist
+	}
+
+	manifests, err := Path()
+	if err != nil {
+		return "", err
+	}
+
+	return filepath.Join(manifests, n.Filepath()), nil
+}
+
+func BlobsPath(digest string) (string, error) {
+	// only accept actual sha256 digests
+	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
+	re := regexp.MustCompile(pattern)
+
+	if digest != "" && !re.MatchString(digest) {
+		return "", ErrInvalidDigestFormat
+	}
+
+	digest = strings.ReplaceAll(digest, ":", "-")
+	path := filepath.Join(envconfig.Models(), "blobs", digest)
+	dirPath := filepath.Dir(path)
+	if digest == "" {
+		dirPath = path
+	}
+
+	if err := os.MkdirAll(dirPath, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PruneDirectory removes empty directories recursively.
+func PruneDirectory(path string) error {
+	info, err := os.Lstat(path)
+	if err != nil {
+		return err
+	}
+
+	if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
+		entries, err := os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		for _, entry := range entries {
+			if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
+				return err
+			}
+		}
+
+		entries, err = os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		if len(entries) > 0 {
+			return nil
+		}
+
+		return os.Remove(path)
+	}
+
+	return nil
+}

middleware/openai.go

@@ -609,3 +609,49 @@ func ImageGenerationsMiddleware() gin.HandlerFunc {
 		c.Next()
 	}
 }
+
+func ImageEditsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ImageEditRequest
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Prompt == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "prompt is required"))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		if req.Image == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "image is required"))
+			return
+		}
+
+		genReq, err := openai.FromImageEditRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &ImageWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}

middleware/openai_test.go

@@ -1112,3 +1112,129 @@ func TestImageWriterResponse(t *testing.T) {
 		t.Errorf("expected image data 'dGVzdC1pbWFnZS1kYXRh', got %s", imageResp.Data[0].B64JSON)
 	}
 }
+
+func TestImageEditsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	// Base64-encoded test image (1x1 pixel PNG)
+	testImage := "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII="
+	decodedImage, _ := base64.StdEncoding.DecodeString("iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=")
+
+	testCases := []testCase{
+		{
+			name: "image edit basic",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+			},
+		},
+		{
+			name: "image edit with size",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `",
+				"size": "512x768"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+				Width:  512,
+				Height: 768,
+			},
+		},
+		{
+			name: "image edit missing prompt",
+			body: `{
+				"model": "test-model",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "prompt is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing model",
+			body: `{
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "model is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing image",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "image is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ImageEditsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Error.Message != "" {
+				var errResp openai.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(tc.err, errResp); diff != "" {
+					t.Fatalf("errors did not match:\n%s", diff)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+			}
+
+			if diff := cmp.Diff(&tc.req, capturedRequest); diff != "" {
+				t.Fatalf("requests did not match:\n%s", diff)
+			}
+		})
+	}
+}

ml/backend.go

@@ -162,6 +162,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 	Conv3D(ctx Context, weight Tensor, c, s0, s1, s2, p0, p1, p2, d0, d1, d2 int) Tensor
+	SSMConv(ctx Context, kernel Tensor) Tensor
 
 	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 

ml/backend/ggml/ggml.go

@@ -1641,6 +1641,13 @@ func (t *Tensor) Conv3D(ctx ml.Context, t2 ml.Tensor, c, s0, s1, s2, p0, p1, p2,
 	return tt
 }
 
+func (t *Tensor) SSMConv(ctx ml.Context, kernel ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_ssm_conv(ctx.(*Context).ctx, t.t, kernel.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
 	return &Tensor{
 		b: t.b,

model/models/lfm2/cache.go

@@ -0,0 +1,410 @@
+package lfm2
+
+import (
+	"slices"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
+)
+
+var _ kvcache.Cache = (*HybridCache)(nil)
+
+// HybridCache stores:
+// - a standard causal KV cache for attention layers
+// - a per-sequence recurrent conv state for shortconv layers
+//
+// Conv state shape (per layer, per sequence): [dConv, hiddenSize] where dConv = L_cache - 1.
+// Stored internally as a tensor of shape [dConv * hiddenSize, maxSlots].
+type HybridCache struct {
+	kv *kvcache.Causal
+
+	backend      ml.Backend
+	dtype        ml.DType
+	maxSequences int
+
+	hiddenSize int
+	dConv      int
+
+	// slot mapping for recurrent state
+	slotForSeq map[int]int
+	refCount   []int
+	freeSlots  []int
+
+	// per-layer conv state buffers (allocated lazily)
+	convCtxs   map[int]ml.Context
+	convStates map[int]ml.Tensor // [dConv*hiddenSize, maxSlots]
+
+	// current forward batch (derived in StartForward)
+	curSeqs       []int
+	curSlots      []int
+	curSlotsInput ml.Tensor
+	curSeqTokens  int
+
+	// track if EnsureWritable has been called for this forward pass
+	writableEnsured bool
+	// track any error from EnsureWritable to propagate later
+	writableError error
+}
+
+func NewHybridCache(shift func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error), hiddenSize, dConv int) *HybridCache {
+	return &HybridCache{
+		kv:         kvcache.NewCausalCache(shift),
+		hiddenSize: hiddenSize,
+		dConv:      dConv,
+		slotForSeq: make(map[int]int),
+		convCtxs:   make(map[int]ml.Context),
+		convStates: make(map[int]ml.Tensor),
+	}
+}
+
+func (c *HybridCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
+	c.backend = backend
+	c.dtype = dtype
+	c.maxSequences = maxSequences
+
+	// initialize slot allocator
+	c.refCount = make([]int, maxSequences)
+	c.freeSlots = c.freeSlots[:0]
+	for i := maxSequences - 1; i >= 0; i-- {
+		c.freeSlots = append(c.freeSlots, i)
+	}
+
+	c.kv.Init(backend, dtype, maxSequences, capacity, maxBatch)
+}
+
+func (c *HybridCache) Close() {
+	for _, ctx := range c.convCtxs {
+		ctx.Close()
+	}
+	c.kv.Close()
+}
+
+func (c *HybridCache) SetConfig(config ml.CacheConfig) {
+	c.kv.SetConfig(config)
+}
+
+func (c *HybridCache) SetLayer(layer int) {
+	c.kv.SetLayer(layer)
+}
+
+func (c *HybridCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
+	return c.kv.Get(ctx)
+}
+
+func (c *HybridCache) Put(ctx ml.Context, key, value ml.Tensor) {
+	c.kv.Put(ctx, key, value)
+}
+
+func (c *HybridCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	if err := c.kv.StartForward(ctx, batch, reserve); err != nil {
+		return err
+	}
+
+	// Derive equal-length sequence layout for shortconv.
+	// LFM2 shortconv assumes tokens form a [seq_tokens, seqs] grid.
+	seqCounts := make(map[int]int)
+	c.curSeqs = c.curSeqs[:0]
+	for _, s := range batch.Sequences {
+		if _, ok := seqCounts[s]; !ok {
+			c.curSeqs = append(c.curSeqs, s)
+		}
+		seqCounts[s]++
+	}
+
+	if len(c.curSeqs) == 0 {
+		return nil
+	}
+
+	nTokens := len(batch.Sequences)
+	nSeqs := len(c.curSeqs)
+	want := nTokens / nSeqs
+	for _, s := range c.curSeqs {
+		if seqCounts[s] != want {
+			return kvcache.ErrNotSupported
+		}
+	}
+
+	c.curSeqTokens = want
+
+	// When reserving memory for estimation, use fake slot assignments
+	// without modifying permanent state (slotForSeq, refCount)
+	if reserve {
+		c.curSlots = c.curSlots[:0]
+		slots := make([]int32, nSeqs)
+		for i := range nSeqs {
+			c.curSlots = append(c.curSlots, i)
+			slots[i] = int32(i)
+		}
+		c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+		return nil
+	}
+
+	// Ensure slots exist for sequences in this batch
+	c.curSlots = c.curSlots[:0]
+	var newSlots []int // track newly allocated slots that need zeroing
+	for _, s := range c.curSeqs {
+		slot, ok := c.slotForSeq[s]
+		if !ok {
+			var err error
+			slot, err = c.allocSlot()
+			if err != nil {
+				return err
+			}
+			c.slotForSeq[s] = slot
+			c.refCount[slot] = 1
+			newSlots = append(newSlots, slot)
+		}
+		c.curSlots = append(c.curSlots, slot)
+	}
+
+	// Zero conv state for newly allocated slots to clear stale data from previous sequences
+	if len(newSlots) > 0 {
+		c.zeroConvSlots(ctx, newSlots)
+	}
+
+	// Create a tensor for the current slots
+	slots := make([]int32, len(c.curSlots))
+	for i, v := range c.curSlots {
+		slots[i] = int32(v)
+	}
+	c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+
+	// Reset writable state for new forward pass
+	c.writableEnsured = false
+	c.writableError = nil
+
+	return nil
+}
+
+func (c *HybridCache) allocSlot() (int, error) {
+	if len(c.freeSlots) == 0 {
+		return 0, kvcache.ErrKvCacheFull
+	}
+	slot := c.freeSlots[len(c.freeSlots)-1]
+	c.freeSlots = c.freeSlots[:len(c.freeSlots)-1]
+	return slot, nil
+}
+
+func (c *HybridCache) freeSlot(slot int) {
+	// Bounds check before freeing
+	if slot >= 0 && slot < c.maxSequences {
+		c.freeSlots = append(c.freeSlots, slot)
+	}
+}
+
+// zeroConvSlots zeros the conv state for the given slots across all layers.
+// This must be called when recycling slots to prevent stale state from affecting new sequences.
+func (c *HybridCache) zeroConvSlots(ctx ml.Context, slots []int) {
+	if len(slots) == 0 || len(c.convStates) == 0 {
+		return
+	}
+
+	// Use input context for creating tensors
+	inputCtx := ctx.Input()
+
+	// Create slot indices tensor
+	slotIndices := make([]int32, len(slots))
+	for i, s := range slots {
+		slotIndices[i] = int32(s)
+	}
+	slotsTensor := inputCtx.FromInts(slotIndices, len(slotIndices))
+
+	// Create zero tensor for the slots (SetRows requires F32 source)
+	zeros := inputCtx.Zeros(ml.DTypeF32, c.dConv*c.hiddenSize, len(slots))
+
+	// Zero each layer's conv state for these slots
+	for _, buf := range c.convStates {
+		ctx.Forward(buf.SetRows(ctx, zeros, slotsTensor))
+	}
+}
+
+// EnsureWritable ensures that sequences in the current batch have private (non-shared) conv slots.
+// Returns an error if slot allocation fails.
+func (c *HybridCache) EnsureWritable(ctx ml.Context) error {
+	for i, seq := range c.curSeqs {
+		slot, ok := c.slotForSeq[seq]
+		if !ok {
+			continue
+		}
+
+		// Bounds check
+		if slot < 0 || slot >= len(c.refCount) {
+			continue
+		}
+
+		if c.refCount[slot] <= 1 {
+			continue
+		}
+
+		newSlot, err := c.allocSlot()
+		if err != nil {
+			return err
+		}
+		c.refCount[slot]--
+		c.refCount[newSlot] = 1
+		c.slotForSeq[seq] = newSlot
+		c.curSlots[i] = newSlot
+
+		// Copy existing conv state for all initialized layers
+		for _, buf := range c.convStates {
+			// buf: [dConv*hiddenSize, maxSlots]
+			src := buf.Rows(ctx, ctx.Input().FromInts([]int32{int32(slot)}, 1))
+			// SetRows requires F32 source
+			srcF32 := src.Cast(ctx, ml.DTypeF32)
+			ctx.Forward(buf.SetRows(ctx, srcF32, ctx.Input().FromInts([]int32{int32(newSlot)}, 1)))
+		}
+	}
+
+	// Rebuild current slots tensor
+	slots := make([]int32, len(c.curSlots))
+	for i, v := range c.curSlots {
+		slots[i] = int32(v)
+	}
+	c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+
+	return nil
+}
+
+func (c *HybridCache) CopyPrefix(srcSeq, dstSeq int, prefixLen int32) {
+	// KV cache shares prefix metadata (no copy) which is correct for prefix reuse.
+	c.kv.CopyPrefix(srcSeq, dstSeq, prefixLen)
+
+	// For shortconv state we implement copy-on-write: dst shares the same slot as src.
+	// On the first write to dst, EnsureWritable will create a private slot.
+	if dstSlot, ok := c.slotForSeq[dstSeq]; ok {
+		// Bounds check before decrementing
+		if dstSlot >= 0 && dstSlot < len(c.refCount) {
+			c.refCount[dstSlot]--
+			if c.refCount[dstSlot] <= 0 {
+				c.refCount[dstSlot] = 0
+				c.freeSlot(dstSlot)
+			}
+		}
+		delete(c.slotForSeq, dstSeq)
+	}
+
+	srcSlot, ok := c.slotForSeq[srcSeq]
+	if !ok {
+		// src may not have a slot yet; dst will allocate on demand
+		return
+	}
+
+	// Bounds check before incrementing
+	if srcSlot >= 0 && srcSlot < len(c.refCount) {
+		c.slotForSeq[dstSeq] = srcSlot
+		c.refCount[srcSlot]++
+	}
+}
+
+func (c *HybridCache) CanResume(seq int, pos int32) bool {
+	return c.kv.CanResume(seq, pos)
+}
+
+func (c *HybridCache) Remove(seq int, beginIndex, endIndex int32) error {
+	if err := c.kv.Remove(seq, beginIndex, endIndex); err != nil {
+		return err
+	}
+
+	// For recurrent state, any removal invalidates the state because
+	// the state at position N depends on all previous positions.
+	// Drop the slot mapping so it resets on next use.
+	slot, ok := c.slotForSeq[seq]
+	if !ok {
+		return nil
+	}
+
+	// Bounds check
+	if slot < 0 || slot >= len(c.refCount) {
+		delete(c.slotForSeq, seq)
+		return nil
+	}
+
+	c.refCount[slot]--
+	if c.refCount[slot] <= 0 {
+		c.refCount[slot] = 0
+		c.freeSlot(slot)
+	}
+	delete(c.slotForSeq, seq)
+
+	return nil
+}
+
+func (c *HybridCache) slotsTensor() ml.Tensor {
+	return c.curSlotsInput
+}
+
+func (c *HybridCache) seqTokens() int {
+	return c.curSeqTokens
+}
+
+func (c *HybridCache) numSeqs() int {
+	return len(c.curSeqs)
+}
+
+func (c *HybridCache) convBuffer(ctx ml.Context, layer int) ml.Tensor {
+	if buf, ok := c.convStates[layer]; ok {
+		return buf
+	}
+
+	if _, ok := c.convCtxs[layer]; !ok {
+		c.convCtxs[layer] = c.backend.NewContextSize(1).Layer(layer)
+	}
+
+	buf := c.convCtxs[layer].Zeros(c.dtype, c.dConv*c.hiddenSize, c.maxSequences)
+	c.convStates[layer] = buf
+	return buf
+}
+
+// ConvState returns the conv state for current batch sequences as shape [dConv, hiddenSize, nSeqs].
+// Returns an error if copy-on-write allocation fails.
+func (c *HybridCache) ConvState(ctx ml.Context, layer int) (ml.Tensor, error) {
+	if !c.writableEnsured {
+		needsWritable := false
+		for _, seq := range c.curSeqs {
+			slot, ok := c.slotForSeq[seq]
+			if !ok {
+				continue
+			}
+			if slot >= 0 && slot < len(c.refCount) && c.refCount[slot] > 1 {
+				needsWritable = true
+				break
+			}
+		}
+
+		if needsWritable {
+			if err := c.EnsureWritable(ctx); err != nil {
+				c.writableError = err
+			}
+		}
+		c.writableEnsured = true
+	}
+
+	if c.writableError != nil {
+		return nil, c.writableError
+	}
+
+	buf := c.convBuffer(ctx, layer)
+	cur := buf.Rows(ctx, c.slotsTensor())
+	return cur.Reshape(ctx, c.dConv, c.hiddenSize, c.numSeqs()), nil
+}
+
+// UpdateConvState writes a new conv state for current batch sequences.
+// newState must have shape [dConv, hiddenSize, nSeqs].
+func (c *HybridCache) UpdateConvState(ctx ml.Context, layer int, newState ml.Tensor) {
+	buf := c.convBuffer(ctx, layer)
+	src := newState.Reshape(ctx, c.dConv*c.hiddenSize, c.numSeqs())
+	// SetRows requires F32 source
+	srcF32 := src.Cast(ctx, ml.DTypeF32)
+	ctx.Forward(buf.SetRows(ctx, srcF32, c.slotsTensor()))
+}
+
+// IsSupportedForBatch returns true if the current batch layout supports shortconv.
+func (c *HybridCache) IsSupportedForBatch() bool {
+	return c.curSeqTokens > 0 && len(c.curSeqs) > 0
+}
+
+// Seqs returns the ordered unique sequences for the current forward pass.
+func (c *HybridCache) Seqs() []int {
+	return slices.Clone(c.curSeqs)
+}

model/models/lfm2/cache_test.go

@@ -0,0 +1,444 @@
+package lfm2
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+)
+
+// TestHybridCache tests verify the slot management logic of HybridCache.
+// These tests focus on the recurrent state slot allocation, reference counting,
+// and copy-on-write semantics without requiring a full ML backend.
+
+// createSlotOnlyCache creates a HybridCache with only the slot management
+// fields initialized. Used to test slot logic in isolation.
+func createSlotOnlyCache(maxSequences int) *HybridCache {
+	return &HybridCache{
+		hiddenSize:   256,
+		dConv:        3,
+		maxSequences: maxSequences,
+		refCount:     make([]int, maxSequences),
+		freeSlots:    initFreeSlots(maxSequences),
+		slotForSeq:   make(map[int]int),
+		convCtxs:     make(map[int]ml.Context),
+		convStates:   make(map[int]ml.Tensor),
+	}
+}
+
+func initFreeSlots(n int) []int {
+	slots := make([]int, 0, n)
+	for i := n - 1; i >= 0; i-- {
+		slots = append(slots, i)
+	}
+	return slots
+}
+
+func TestHybridCache_SlotAllocation(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Verify initial state
+	if len(cache.freeSlots) != 4 {
+		t.Errorf("expected 4 free slots, got %d", len(cache.freeSlots))
+	}
+
+	// Allocate all slots
+	for range 4 {
+		slot, err := cache.allocSlot()
+		if err != nil {
+			t.Fatalf("allocSlot failed: %v", err)
+		}
+		cache.refCount[slot] = 1
+	}
+
+	// Should be full now
+	if len(cache.freeSlots) != 0 {
+		t.Errorf("expected 0 free slots, got %d", len(cache.freeSlots))
+	}
+
+	// Trying to allocate another should fail
+	_, err := cache.allocSlot()
+	if err != kvcache.ErrKvCacheFull {
+		t.Errorf("expected ErrKvCacheFull, got %v", err)
+	}
+}
+
+func TestHybridCache_SlotReuse(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate a slot
+	slot1, _ := cache.allocSlot()
+	cache.refCount[slot1] = 1
+
+	// Free it
+	cache.refCount[slot1] = 0
+	cache.freeSlot(slot1)
+
+	// Allocate again - should get the same slot back (LIFO)
+	slot2, _ := cache.allocSlot()
+	if slot2 != slot1 {
+		t.Errorf("expected slot %d to be reused, got %d", slot1, slot2)
+	}
+}
+
+func TestHybridCache_SlotRefCounting_ShareSlot(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Simulate sharing slot with seq 2 (copy-on-write style)
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Should share the same slot
+	if cache.slotForSeq[2] != slot1 {
+		t.Errorf("expected seq 2 to share slot %d, got %d", slot1, cache.slotForSeq[2])
+	}
+
+	// Ref count should be 2
+	if cache.refCount[slot1] != 2 {
+		t.Errorf("expected refCount 2, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_SlotRefCounting_DecRef(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Share with seq 2
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Unshare seq 2
+	cache.refCount[slot1]--
+	delete(cache.slotForSeq, 2)
+
+	// Ref count should be back to 1
+	if cache.refCount[slot1] != 1 {
+		t.Errorf("expected refCount 1 after unshare, got %d", cache.refCount[slot1])
+	}
+
+	// Seq 2 should no longer have a slot
+	if _, ok := cache.slotForSeq[2]; ok {
+		t.Error("seq 2 should not have a slot after unshare")
+	}
+}
+
+func TestHybridCache_SlotFreeWhenUnused(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	initialFreeSlots := len(cache.freeSlots)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Free the slot when refCount drops to 0
+	cache.refCount[slot1]--
+	if cache.refCount[slot1] <= 0 {
+		cache.refCount[slot1] = 0
+		cache.freeSlot(slot1)
+	}
+	delete(cache.slotForSeq, 1)
+
+	// Slot should be freed
+	if len(cache.freeSlots) != initialFreeSlots {
+		t.Errorf("expected %d free slots, got %d", initialFreeSlots, len(cache.freeSlots))
+	}
+
+	// Ref count should be 0
+	if cache.refCount[slot1] != 0 {
+		t.Errorf("expected refCount 0, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_SlotOverwrite(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slots for seq 1 and seq 2
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	slot2, _ := cache.allocSlot()
+	cache.slotForSeq[2] = slot2
+	cache.refCount[slot2] = 1
+
+	initialFreeSlots := len(cache.freeSlots)
+
+	// Simulate overwriting seq 2's slot with slot1 (sharing)
+	// First free the old slot
+	cache.refCount[slot2]--
+	if cache.refCount[slot2] <= 0 {
+		cache.refCount[slot2] = 0
+		cache.freeSlot(slot2)
+	}
+	// Then share slot1
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Seq 2 should now share slot1
+	if cache.slotForSeq[2] != slot1 {
+		t.Errorf("expected seq 2 to share slot %d, got %d", slot1, cache.slotForSeq[2])
+	}
+
+	// Old slot2 should be freed
+	if len(cache.freeSlots) != initialFreeSlots+1 {
+		t.Errorf("expected %d free slots, got %d", initialFreeSlots+1, len(cache.freeSlots))
+	}
+}
+
+func TestHybridCache_BoundsChecking(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Test freeing invalid slot (should not panic)
+	cache.freeSlot(-1)
+	cache.freeSlot(100) // out of bounds
+
+	// freeSlot does bounds checking, so invalid slots should be ignored
+	if len(cache.freeSlots) != 4 {
+		t.Errorf("invalid slots should not affect free list, got %d slots", len(cache.freeSlots))
+	}
+}
+
+func TestHybridCache_MultipleSequences_RefCounting(t *testing.T) {
+	cache := createSlotOnlyCache(8)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Fork to seq 2, 3, 4 (all share slot1)
+	for _, seq := range []int{2, 3, 4} {
+		cache.slotForSeq[seq] = slot1
+		cache.refCount[slot1]++
+	}
+
+	// Ref count should be 4
+	if cache.refCount[slot1] != 4 {
+		t.Errorf("expected refCount 4, got %d", cache.refCount[slot1])
+	}
+
+	// Remove seq 2, 3
+	for _, seq := range []int{2, 3} {
+		delete(cache.slotForSeq, seq)
+		cache.refCount[slot1]--
+	}
+
+	if cache.refCount[slot1] != 2 {
+		t.Errorf("expected refCount 2, got %d", cache.refCount[slot1])
+	}
+
+	// Slot should still be allocated (not in free list)
+	found := false
+	for _, s := range cache.freeSlots {
+		if s == slot1 {
+			found = true
+			break
+		}
+	}
+	if found {
+		t.Error("slot1 should not be in free list yet")
+	}
+
+	// Remove remaining sequences
+	for _, seq := range []int{1, 4} {
+		delete(cache.slotForSeq, seq)
+		cache.refCount[slot1]--
+	}
+
+	if cache.refCount[slot1] != 0 {
+		t.Errorf("expected refCount 0, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_ChainedSharing(t *testing.T) {
+	cache := createSlotOnlyCache(8)
+
+	// Create seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Share 1 -> 2
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Share 2 -> 3 (should still share slot1)
+	cache.slotForSeq[3] = cache.slotForSeq[2] // which is slot1
+	cache.refCount[slot1]++
+
+	// All should share slot1
+	if cache.slotForSeq[1] != slot1 || cache.slotForSeq[2] != slot1 || cache.slotForSeq[3] != slot1 {
+		t.Error("all sequences should share slot1")
+	}
+
+	if cache.refCount[slot1] != 3 {
+		t.Errorf("expected refCount 3, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_CacheParameters(t *testing.T) {
+	cache := NewHybridCache(nil, 512, 5) // hiddenSize=512, dConv=5
+
+	if cache.hiddenSize != 512 {
+		t.Errorf("expected hiddenSize 512, got %d", cache.hiddenSize)
+	}
+	if cache.dConv != 5 {
+		t.Errorf("expected dConv 5, got %d", cache.dConv)
+	}
+}
+
+func TestHybridCache_NumSeqs(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Initially no sequences
+	if cache.numSeqs() != 0 {
+		t.Errorf("expected 0 seqs, got %d", cache.numSeqs())
+	}
+
+	// Manually set up current batch state
+	cache.curSeqs = []int{1, 2, 3}
+
+	if cache.numSeqs() != 3 {
+		t.Errorf("expected 3 seqs, got %d", cache.numSeqs())
+	}
+}
+
+func TestHybridCache_SeqTokens(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Initially 0
+	if cache.seqTokens() != 0 {
+		t.Errorf("expected 0 seqTokens, got %d", cache.seqTokens())
+	}
+
+	// Manually set up current batch state
+	cache.curSeqTokens = 16
+
+	if cache.seqTokens() != 16 {
+		t.Errorf("expected 16 seqTokens, got %d", cache.seqTokens())
+	}
+}
+
+// Test that Seqs returns a clone of curSeqs
+func TestHybridCache_Seqs_ReturnsClone(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	cache.curSeqs = []int{1, 2, 3}
+
+	seqs := cache.Seqs()
+
+	// Modify returned slice
+	seqs[0] = 999
+
+	// Original should be unchanged
+	if cache.curSeqs[0] != 1 {
+		t.Error("Seqs should return a clone, not the original slice")
+	}
+}
+
+func TestHybridCache_IsSupportedForBatch(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Initially not supported (no batch set up)
+	if cache.IsSupportedForBatch() {
+		t.Error("expected IsSupportedForBatch to be false initially")
+	}
+
+	// Set up a valid batch
+	cache.curSeqTokens = 1
+	cache.curSeqs = []int{1}
+
+	if !cache.IsSupportedForBatch() {
+		t.Error("expected IsSupportedForBatch to be true with valid batch")
+	}
+}
+
+func TestHybridCache_ZeroConvSlots_EmptyInputs(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// zeroConvSlots should handle empty slots without panicking
+	cache.zeroConvSlots(nil, nil)
+	cache.zeroConvSlots(nil, []int{})
+
+	// zeroConvSlots should handle empty convStates without panicking
+	cache.zeroConvSlots(nil, []int{0, 1, 2})
+}
+
+func TestHybridCache_SlotRecycling_TracksNewSlots(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Free the slot (simulating sequence removal)
+	cache.refCount[slot1]--
+	cache.freeSlot(slot1)
+	delete(cache.slotForSeq, 1)
+
+	// Verify slot is in free list
+	if len(cache.freeSlots) != 4 {
+		t.Errorf("expected 4 free slots after freeing, got %d", len(cache.freeSlots))
+	}
+
+	// Allocate for new seq 2 - should get recycled slot
+	slot2, _ := cache.allocSlot()
+	if slot2 != slot1 {
+		t.Errorf("expected recycled slot %d, got %d", slot1, slot2)
+	}
+
+	// This recycled slot would need zeroing in the real implementation
+	// The actual zeroing is tested via integration tests since it requires ML context
+}
+
+func TestHybridCache_NewSequence_GetsTrackedForZeroing(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Simulate the slot allocation flow from StartForward
+	// When a sequence doesn't have a slot, it gets allocated and tracked as "new"
+
+	newSlots := []int{}
+
+	// Seq 1 doesn't have a slot - allocate and track
+	seq := 1
+	if _, ok := cache.slotForSeq[seq]; !ok {
+		slot, err := cache.allocSlot()
+		if err != nil {
+			t.Fatalf("allocSlot failed: %v", err)
+		}
+		cache.slotForSeq[seq] = slot
+		cache.refCount[slot] = 1
+		newSlots = append(newSlots, slot)
+	}
+
+	// Verify newSlots contains the allocated slot
+	if len(newSlots) != 1 {
+		t.Errorf("expected 1 new slot, got %d", len(newSlots))
+	}
+
+	// Seq 1 already has a slot - should NOT be tracked as new
+	newSlots2 := []int{}
+	if _, ok := cache.slotForSeq[seq]; !ok {
+		slot, _ := cache.allocSlot()
+		cache.slotForSeq[seq] = slot
+		cache.refCount[slot] = 1
+		newSlots2 = append(newSlots2, slot)
+	}
+
+	// Verify no new slots for existing sequence
+	if len(newSlots2) != 0 {
+		t.Errorf("expected 0 new slots for existing sequence, got %d", len(newSlots2))
+	}
+}

model/models/lfm2/model.go

@@ -0,0 +1,253 @@
+package lfm2
+
+import (
+	"cmp"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	hiddenSize       int
+	headDim, ropeDim int
+
+	eps, ropeBase, ropeScale float32
+
+	ropeType              string
+	originalContextLength int
+
+	// per-layer head counts (LFM2 alternates attention and recurrent layers)
+	numHeadsByLayer   []int
+	numKVHeadsByLayer []int
+}
+
+func (o Options) headDimValue() int {
+	// Head dim is shared across layers; fall back to first attention layer head count.
+	for _, h := range o.numHeadsByLayer {
+		if h > 0 {
+			return cmp.Or(o.headDim, o.hiddenSize/h)
+		}
+	}
+	return cmp.Or(o.headDim, o.hiddenSize)
+}
+
+func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
+	opts := []func(*rope.Options){rope.WithTypeNeoX()}
+	if o.ropeType == "yarn" {
+		attnFactor := float32(1.0 / (1.0 + 0.1*math.Log(float64(o.ropeScale))))
+		opts = append(opts,
+			rope.WithOriginalContextLength(o.originalContextLength),
+			rope.WithExtrapolationFactor(1.),
+			rope.WithAttentionFactor(attnFactor),
+		)
+	}
+
+	headCount := 1
+	for _, h := range o.numHeadsByLayer {
+		if h > 0 {
+			headCount = h
+			break
+		}
+	}
+	return nn.RoPE(ctx, states, positions, cmp.Or(o.ropeDim, o.headDim, o.hiddenSize/headCount), o.ropeBase, 1./o.ropeScale, opts...)
+}
+
+type Model struct {
+	model.Base
+	model.TextProcessor
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm,alt:token_embd_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
+
+	Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	if c.Uint("expert_count") > 0 {
+		return nil, model.ErrUnsupportedModel
+	}
+
+	if c.String("tokenizer.ggml.model") != "gpt2" {
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	vocabulary := model.Vocabulary{
+		Values: c.Strings("tokenizer.ggml.tokens"),
+		Scores: c.Floats("tokenizer.ggml.scores"),
+		Types:  c.Ints("tokenizer.ggml.token_type"),
+		Merges: c.Strings("tokenizer.ggml.merges"),
+		AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+		BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+		AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+		EOS: append(
+			[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+			c.Ints("tokenizer.ggml.eos_token_ids")...,
+		),
+	}
+
+	var pretokenizers []string
+	switch c.String("tokenizer.ggml.pre") {
+	case "default":
+		// use default BPE pretokenizer
+	default:
+		// llama-bpe style (default for LFM2)
+		pretokenizers = []string{
+			`(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
+		}
+	}
+
+	m := Model{
+		TextProcessor: model.NewBytePairEncoding(&vocabulary, pretokenizers...),
+		Layers:        make([]Layer, c.Uint("block_count")),
+		Options: Options{
+			hiddenSize:            int(c.Uint("embedding_length")),
+			headDim:               int(c.Uint("attention.key_length")),
+			ropeDim:               int(c.Uint("rope.dimension_count")),
+			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
+			ropeType:              c.String("rope.scaling.type"),
+			ropeBase:              c.Float("rope.freq_base"),
+			ropeScale:             c.Float("rope.scaling.factor", 1),
+			originalContextLength: int(c.Uint("rope.scaling.original_context_length")),
+		},
+	}
+
+	type headCounts interface {
+		HeadCount() []uint64
+		HeadCountKV() []uint64
+	}
+	hc, ok := c.(headCounts)
+	if !ok {
+		return nil, model.ErrUnsupportedModel
+	}
+
+	headCount := hc.HeadCount()
+	headCountKV := hc.HeadCountKV()
+
+	m.numHeadsByLayer = make([]int, len(m.Layers))
+	m.numKVHeadsByLayer = make([]int, len(m.Layers))
+	for i := range m.Layers {
+		m.numHeadsByLayer[i] = int(headCount[i])
+		m.numKVHeadsByLayer[i] = int(headCountKV[i])
+
+		if m.numKVHeadsByLayer[i] == 0 {
+			m.Layers[i].Operator = &ShortConv{}
+		} else {
+			m.Layers[i].Operator = &Attention{}
+		}
+	}
+
+	lCache := int(c.Uint("shortconv.l_cache"))
+	dConv := max(0, lCache-1)
+	m.Cache = NewHybridCache(m.Shift, m.hiddenSize, dConv)
+	return &m, nil
+}
+
+type Operator interface {
+	Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor
+}
+
+type Attention struct {
+	Query     *nn.Linear  `gguf:"attn_q"`
+	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
+	Key       *nn.Linear  `gguf:"attn_k"`
+	KeyNorm   *nn.RMSNorm `gguf:"attn_k_norm"`
+	Value     *nn.Linear  `gguf:"attn_v"`
+	Output    *nn.Linear  `gguf:"attn_output,alt:attn_out"`
+}
+
+func (sa *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor {
+	batchSize := hiddenStates.Dim(1)
+	headDim := opts.headDimValue()
+	numHeads := opts.numHeadsByLayer[layer]
+	numKVHeads := opts.numKVHeadsByLayer[layer]
+
+	query := sa.Query.Forward(ctx, hiddenStates)
+	key := sa.Key.Forward(ctx, hiddenStates)
+	value := sa.Value.Forward(ctx, hiddenStates)
+
+	query = query.Reshape(ctx, headDim, numHeads, batchSize)
+	key = key.Reshape(ctx, headDim, numKVHeads, batchSize)
+	value = value.Reshape(ctx, headDim, numKVHeads, batchSize)
+
+	query = sa.QueryNorm.Forward(ctx, query, opts.eps)
+	key = sa.KeyNorm.Forward(ctx, key, opts.eps)
+
+	query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
+	key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
+
+	attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(headDim)), cache)
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)
+	return sa.Output.Forward(ctx, attention)
+}
+
+type MLP struct {
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+	Gate *nn.Linear `gguf:"ffn_gate"`
+}
+
+func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	return mlp.Down.Forward(ctx, hiddenState)
+}
+
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	Operator      Operator
+	MLPNorm       *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP           *MLP
+}
+
+func (l *Layer) Forward(ctx ml.Context, layer int, hiddenState, positions, outputs ml.Tensor, cache *HybridCache, opts *Options) ml.Tensor {
+	residual := hiddenState
+
+	hiddenState = l.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.Operator.Forward(ctx, hiddenState, positions, cache, layer, opts)
+
+	if outputs != nil {
+		hiddenState = hiddenState.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenState = hiddenState.Add(ctx, residual)
+	residual = hiddenState
+
+	hiddenState = l.MLPNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.MLP.Forward(ctx, hiddenState, opts)
+	return hiddenState.Add(ctx, residual)
+}
+
+func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
+}
+
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+
+	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = batch.Outputs
+		}
+
+		hiddenState = layer.Forward(ctx, i, hiddenState, positions, outputs, m.Cache.(*HybridCache), &m.Options)
+	}
+
+	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
+	return m.Output.Forward(ctx, hiddenState), nil
+}
+
+func init() {
+	model.Register("lfm2", New)
+}

model/models/lfm2/shortconv.go

@@ -0,0 +1,50 @@
+package lfm2
+
+import (
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+)
+
+type shortConvKernel struct {
+	Weight ml.Tensor `gguf:"weight"`
+}
+
+// ShortConv implements the LFM2 short-convolution block (GGML_OP_SSM_CONV) with a recurrent
+// state stored in the HybridCache.
+type ShortConv struct {
+	Conv    *shortConvKernel `gguf:"shortconv.conv"`
+	InProj  *nn.Linear       `gguf:"shortconv.in_proj"`
+	OutProj *nn.Linear       `gguf:"shortconv.out_proj"`
+}
+
+func (sc *ShortConv) Forward(ctx ml.Context, hiddenStates ml.Tensor, _ ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor {
+	nSeqs := cache.numSeqs()
+	seqTokens := cache.seqTokens()
+	hiddenSize := hiddenStates.Dim(0)
+	if nSeqs <= 0 || seqTokens <= 0 || hiddenStates.Dim(1) != nSeqs*seqTokens {
+		panic("lfm2: unsupported batch layout for shortconv")
+	}
+
+	bcx := sc.InProj.Forward(ctx, hiddenStates).Reshape(ctx, 3*hiddenSize, seqTokens, nSeqs)
+
+	elementSize := bcx.Stride(0)
+	b := bcx.View(ctx, 0*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
+	c := bcx.View(ctx, 1*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
+	x := bcx.View(ctx, 2*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
+
+	bx := b.Mul(ctx, x).Permute(ctx, 1, 0, 2, 3)
+
+	state, err := cache.ConvState(ctx, layer)
+	if err != nil {
+		panic("lfm2: failed to get conv state: " + err.Error())
+	}
+	sx := state.Concat(ctx, bx, 0)
+
+	convOut := sx.SSMConv(ctx, sc.Conv.Weight)
+	y := c.Mul(ctx, convOut)
+
+	dConv := sx.Dim(0) - seqTokens
+	cache.UpdateConvState(ctx, layer, sx.Slice(ctx, 0, sx.Dim(0)-dConv, sx.Dim(0), 1))
+
+	return sc.OutProj.Forward(ctx, y.Reshape(ctx, hiddenSize, seqTokens*nSeqs))
+}

model/models/models.go

@@ -9,6 +9,7 @@ import (
 	_ "github.com/ollama/ollama/model/models/gemma3n"
 	_ "github.com/ollama/ollama/model/models/glm4moelite"
 	_ "github.com/ollama/ollama/model/models/gptoss"
+	_ "github.com/ollama/ollama/model/models/lfm2"
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/llama4"
 	_ "github.com/ollama/ollama/model/models/mistral3"

model/parsers/lfm2.go

@@ -0,0 +1,498 @@
+package parsers
+
+import (
+	"encoding/json"
+	"errors"
+	"log/slog"
+	"strconv"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+)
+
+type LFM2ParserState int
+
+const (
+	LFM2CollectingThinking LFM2ParserState = iota
+	LFM2CollectingContent
+	LFM2CollectingToolCalls
+)
+
+const (
+	lfm2ThinkingOpenTag  = "<think>"
+	lfm2ThinkingCloseTag = "</think>"
+	lfm2ToolCallStartTag = "<|tool_call_start|>"
+	lfm2ToolCallEndTag   = "<|tool_call_end|>"
+)
+
+type LFM2Parser struct {
+	state                    LFM2ParserState
+	buffer                   strings.Builder
+	hasThinkingSupport       bool
+	needsThinkingLeadingTrim bool // trim leading whitespace after <think> tag
+	needsContentLeadingTrim  bool // trim leading whitespace after </think> tag
+}
+
+func (p *LFM2Parser) HasToolSupport() bool {
+	return true
+}
+
+func (p *LFM2Parser) HasThinkingSupport() bool {
+	return p.hasThinkingSupport
+}
+
+func (p *LFM2Parser) setInitialState(lastMessage *api.Message, thinkValue *api.ThinkValue) {
+	prefill := lastMessage != nil && lastMessage.Role == "assistant"
+
+	// Check both model capability AND request preference
+	thinkingEnabled := p.HasThinkingSupport() && (thinkValue != nil && thinkValue.Bool())
+
+	if !thinkingEnabled {
+		p.state = LFM2CollectingContent
+		return
+	}
+
+	if prefill && lastMessage.Content != "" {
+		p.state = LFM2CollectingContent
+		return
+	}
+
+	p.state = LFM2CollectingThinking
+	p.needsThinkingLeadingTrim = true
+}
+
+func (p *LFM2Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.setInitialState(lastMessage, thinkValue)
+	return tools
+}
+
+type lfm2Event interface {
+	isLFM2Event()
+}
+
+type lfm2EventThinkingContent struct {
+	content string
+}
+
+type lfm2EventContent struct {
+	content string
+}
+
+type lfm2EventToolCall struct {
+	toolCall api.ToolCall
+}
+
+func (lfm2EventThinkingContent) isLFM2Event() {}
+func (lfm2EventContent) isLFM2Event()         {}
+func (lfm2EventToolCall) isLFM2Event()        {}
+
+func (p *LFM2Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+	events := p.parseEvents()
+
+	var toolCalls []api.ToolCall
+	var contentSb strings.Builder
+	var thinkingSb strings.Builder
+	for _, event := range events {
+		switch event := event.(type) {
+		case lfm2EventToolCall:
+			toolCalls = append(toolCalls, event.toolCall)
+		case lfm2EventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case lfm2EventContent:
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *LFM2Parser) parseEvents() []lfm2Event {
+	var all []lfm2Event
+
+	keepLooping := true
+	for keepLooping {
+		var events []lfm2Event
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	return all
+}
+
+func (p *LFM2Parser) eat() ([]lfm2Event, bool) {
+	var events []lfm2Event
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return events, false
+	}
+
+	switch p.state {
+	case LFM2CollectingThinking:
+		// Strip opening <think> tag if present
+		if strings.HasPrefix(bufStr, lfm2ThinkingOpenTag) {
+			bufStr = bufStr[len(lfm2ThinkingOpenTag):]
+			p.needsThinkingLeadingTrim = true
+			p.buffer.Reset()
+			p.buffer.WriteString(bufStr)
+		}
+
+		// Trim leading whitespace after <think> tag (may span multiple chunks)
+		if p.needsThinkingLeadingTrim {
+			if trimmed := strings.TrimLeftFunc(bufStr, unicode.IsSpace); trimmed != bufStr {
+				bufStr = trimmed
+				p.buffer.Reset()
+				p.buffer.WriteString(bufStr)
+			}
+			// Clear flag once we have non-whitespace content or buffer is empty
+			if len(bufStr) > 0 {
+				p.needsThinkingLeadingTrim = false
+			}
+		}
+
+		if strings.Contains(bufStr, lfm2ThinkingCloseTag) { // thinking[</think>] -> content
+			split := strings.SplitN(bufStr, lfm2ThinkingCloseTag, 2)
+			thinking := split[0]
+			thinking = strings.TrimRightFunc(thinking, unicode.IsSpace)
+
+			remaining := split[1]
+			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = LFM2CollectingContent
+			p.needsThinkingLeadingTrim = false
+			// Set flag to trim any additional whitespace that may arrive in later chunks
+			p.needsContentLeadingTrim = len(remaining) == 0
+
+			if len(thinking) > 0 {
+				events = append(events, lfm2EventThinkingContent{content: thinking})
+			}
+			return events, true
+		} else if overlapLen := overlap(bufStr, lfm2ThinkingCloseTag); overlapLen > 0 { // partial </think>
+			beforePartialTag := bufStr[:len(bufStr)-overlapLen]
+			trailingLen := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingLen
+
+			unambiguous := bufStr[:ambiguousStart]
+			ambiguous := bufStr[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, lfm2EventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		} else { // otherwise its thinking content
+			whitespaceLen := trailingWhitespaceLen(bufStr)
+			ambiguousStart := len(bufStr) - whitespaceLen
+
+			unambiguous := bufStr[:ambiguousStart]
+			ambiguous := bufStr[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, lfm2EventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case LFM2CollectingContent:
+		// Trim leading whitespace after </think> tag (may span multiple chunks)
+		if p.needsContentLeadingTrim {
+			if trimmed := strings.TrimLeftFunc(bufStr, unicode.IsSpace); trimmed != bufStr {
+				bufStr = trimmed
+				p.buffer.Reset()
+				p.buffer.WriteString(bufStr)
+			}
+			// Clear flag once we have non-whitespace content
+			if len(bufStr) > 0 {
+				p.needsContentLeadingTrim = false
+			}
+		}
+
+		if strings.Contains(bufStr, lfm2ToolCallStartTag) { // content[<|tool_call_start|>] -> tool calls
+			split := strings.SplitN(bufStr, lfm2ToolCallStartTag, 2)
+			contentBefore := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = LFM2CollectingToolCalls
+
+			if len(contentBefore) > 0 {
+				events = append(events, lfm2EventContent{content: contentBefore})
+			}
+			return events, true
+		} else { // otherwise its content
+			p.buffer.Reset()
+			if len(bufStr) > 0 {
+				events = append(events, lfm2EventContent{content: bufStr})
+			}
+			return events, false
+		}
+
+	case LFM2CollectingToolCalls:
+		// Look for complete tool call JSON between tags
+		if idx := strings.Index(bufStr, lfm2ToolCallEndTag); idx != -1 {
+			toolCallContent := bufStr[:idx]
+
+			if toolCalls, err := p.parseToolCallsContent(toolCallContent); err == nil && len(toolCalls) > 0 {
+				remaining := bufStr[idx+len(lfm2ToolCallEndTag):]
+
+				// Check if there's another tool call
+				if strings.HasPrefix(remaining, lfm2ToolCallStartTag) {
+					remaining = remaining[len(lfm2ToolCallStartTag):]
+				} else {
+					// No more tool calls, go back to content
+					remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+					p.state = LFM2CollectingContent
+				}
+
+				p.buffer.Reset()
+				p.buffer.WriteString(remaining)
+
+				for _, tc := range toolCalls {
+					events = append(events, lfm2EventToolCall{toolCall: tc})
+				}
+				return events, true
+			} else if err != nil {
+				slog.Warn("lfm2 tool call parsing failed", "error", err, "content", toolCallContent)
+			}
+		}
+
+		return events, false
+	}
+
+	return events, false
+}
+
+// parseToolCallsContent parses one or more tool calls from content
+// Supports JSON format and Python-style format including multiple calls: [func1(...),func2(...)]
+func (p *LFM2Parser) parseToolCallsContent(content string) ([]api.ToolCall, error) {
+	content = strings.TrimSpace(content)
+
+	// Try JSON format first: {"name": "func", "arguments": {...}}
+	var parsed struct {
+		Name      string          `json:"name"`
+		Arguments json.RawMessage `json:"arguments"`
+	}
+
+	if err := json.Unmarshal([]byte(content), &parsed); err == nil && parsed.Name != "" {
+		var args api.ToolCallFunctionArguments
+		if len(parsed.Arguments) > 0 {
+			if err := json.Unmarshal(parsed.Arguments, &args); err != nil {
+				return nil, err
+			}
+		} else {
+			args = api.NewToolCallFunctionArguments()
+		}
+
+		return []api.ToolCall{{
+			Function: api.ToolCallFunction{
+				Name:      parsed.Name,
+				Arguments: args,
+			},
+		}}, nil
+	}
+
+	// Try Python-style format: [func(arg1='val1'),func2(arg2='val2')] or func(arg1='val1')
+	return p.parsePythonStyleToolCalls(content)
+}
+
+// parsePythonStyleToolCalls parses one or more Python-style tool calls
+// Examples: [bash(command='ls'),bash(command='pwd')] or bash(command='ls')
+func (p *LFM2Parser) parsePythonStyleToolCalls(content string) ([]api.ToolCall, error) {
+	content = strings.TrimSpace(content)
+
+	// Strip outer brackets if present: [func(...)] -> func(...)
+	if strings.HasPrefix(content, "[") && strings.HasSuffix(content, "]") {
+		content = content[1 : len(content)-1]
+	}
+
+	var toolCalls []api.ToolCall
+
+	// Parse multiple function calls separated by commas at the top level
+	for len(content) > 0 {
+		content = strings.TrimSpace(content)
+		if content == "" {
+			break
+		}
+
+		// Skip leading comma from previous iteration
+		if strings.HasPrefix(content, ",") {
+			content = strings.TrimSpace(content[1:])
+			if content == "" {
+				break
+			}
+		}
+
+		// Find function name
+		parenIdx := strings.Index(content, "(")
+		if parenIdx == -1 {
+			return nil, errors.New("invalid tool call: no opening parenthesis")
+		}
+
+		funcName := strings.TrimSpace(content[:parenIdx])
+		if funcName == "" {
+			return nil, errors.New("invalid tool call: empty function name")
+		}
+
+		// Find matching closing parenthesis
+		closeIdx := findMatchingParen(content, parenIdx)
+		if closeIdx == -1 {
+			return nil, errors.New("invalid tool call: no matching closing parenthesis")
+		}
+
+		argsStr := content[parenIdx+1 : closeIdx]
+		args := api.NewToolCallFunctionArguments()
+
+		if argsStr != "" {
+			if err := parsePythonArgs(argsStr, &args); err != nil {
+				return nil, err
+			}
+		}
+
+		toolCalls = append(toolCalls, api.ToolCall{
+			Function: api.ToolCallFunction{
+				Name:      funcName,
+				Arguments: args,
+			},
+		})
+
+		// Move past this function call
+		content = content[closeIdx+1:]
+	}
+
+	if len(toolCalls) == 0 {
+		return nil, errors.New("no tool calls found")
+	}
+
+	return toolCalls, nil
+}
+
+// findMatchingParen finds the index of the closing parenthesis matching the one at openIdx
+// Returns -1 if not found. Handles nested parentheses and quoted strings.
+func findMatchingParen(s string, openIdx int) int {
+	depth := 1
+	i := openIdx + 1
+	for i < len(s) && depth > 0 {
+		switch s[i] {
+		case '(':
+			depth++
+		case ')':
+			depth--
+			if depth == 0 {
+				return i
+			}
+		case '\'', '"':
+			// Skip quoted string
+			quote := s[i]
+			i++
+			for i < len(s) && s[i] != quote {
+				if s[i] == '\\' && i+1 < len(s) {
+					i++ // skip escaped char
+				}
+				i++
+			}
+		}
+		i++
+	}
+	return -1
+}
+
+// parseToolCallContent parses a single tool call (for backward compatibility with tests)
+func (p *LFM2Parser) parseToolCallContent(content string) (api.ToolCall, error) {
+	calls, err := p.parseToolCallsContent(content)
+	if err != nil {
+		return api.ToolCall{}, err
+	}
+	if len(calls) == 0 {
+		return api.ToolCall{}, errors.New("no tool call found")
+	}
+	return calls[0], nil
+}
+
+// parsePythonArgs parses Python-style keyword arguments: key='value', key2="value2"
+func parsePythonArgs(argsStr string, args *api.ToolCallFunctionArguments) error {
+	// Simple state machine to parse key='value' pairs
+	// Handles: command='ls', flag="-la", count=42, enabled=true
+	var key string
+	i := 0
+
+	for i < len(argsStr) {
+		// Skip whitespace
+		for i < len(argsStr) && (argsStr[i] == ' ' || argsStr[i] == '\t' || argsStr[i] == '\n') {
+			i++
+		}
+		if i >= len(argsStr) {
+			break
+		}
+
+		// Parse key
+		keyStart := i
+		for i < len(argsStr) && argsStr[i] != '=' && argsStr[i] != ',' {
+			i++
+		}
+		if i >= len(argsStr) || argsStr[i] != '=' {
+			return errors.New("invalid argument: expected '='")
+		}
+		key = strings.TrimSpace(argsStr[keyStart:i])
+		i++ // skip '='
+
+		// Skip whitespace after =
+		for i < len(argsStr) && (argsStr[i] == ' ' || argsStr[i] == '\t') {
+			i++
+		}
+
+		// Parse value
+		var value string
+		if i < len(argsStr) && (argsStr[i] == '\'' || argsStr[i] == '"') {
+			// Quoted string
+			quote := argsStr[i]
+			i++
+			valueStart := i
+			for i < len(argsStr) && argsStr[i] != quote {
+				if argsStr[i] == '\\' && i+1 < len(argsStr) {
+					i += 2 // skip escaped char
+				} else {
+					i++
+				}
+			}
+			value = argsStr[valueStart:i]
+			if i < len(argsStr) {
+				i++ // skip closing quote
+			}
+			args.Set(key, value)
+		} else {
+			// Unquoted value (number, bool, etc)
+			valueStart := i
+			for i < len(argsStr) && argsStr[i] != ',' {
+				i++
+			}
+			value = strings.TrimSpace(argsStr[valueStart:i])
+
+			// Try to parse as number or bool
+			if v, err := strconv.ParseInt(value, 10, 64); err == nil {
+				args.Set(key, v)
+			} else if v, err := strconv.ParseFloat(value, 64); err == nil {
+				args.Set(key, v)
+			} else if value == "true" {
+				args.Set(key, true)
+			} else if value == "false" {
+				args.Set(key, false)
+			} else {
+				args.Set(key, value)
+			}
+		}
+
+		// Skip comma and whitespace
+		for i < len(argsStr) && (argsStr[i] == ',' || argsStr[i] == ' ' || argsStr[i] == '\t' || argsStr[i] == '\n') {
+			i++
+		}
+	}
+
+	return nil
+}

model/parsers/parsers.go

@@ -70,6 +70,10 @@ func ParserForName(name string) Parser {
 		return &FunctionGemmaParser{}
 	case "glm-4.7":
 		return &GLM47Parser{}
+	case "lfm2":
+		return &LFM2Parser{hasThinkingSupport: false}
+	case "lfm2-thinking":
+		return &LFM2Parser{hasThinkingSupport: true}
 	default:
 		return nil
 	}

model/renderers/lfm2.go

@@ -0,0 +1,144 @@
+package renderers
+
+import (
+	"encoding/json"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+type LFM2Renderer struct {
+	IsThinking bool
+}
+
+func (r *LFM2Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	// Note: BOS token is added by the tokenizer (add_bos_token: true), not the renderer
+
+	// Extract first system message if present (to combine with tools)
+	var firstSystemContent string
+	startIdx := 0
+	if len(messages) > 0 && messages[0].Role == "system" {
+		firstSystemContent = messages[0].Content
+		startIdx = 1
+	}
+
+	// Append tools to first system content
+	if len(tools) > 0 {
+		if firstSystemContent != "" {
+			firstSystemContent += "\n"
+		}
+		firstSystemContent += "List of tools: ["
+		for i, tool := range tools {
+			toolJSON, err := json.Marshal(tool)
+			if err != nil {
+				return "", err
+			}
+			firstSystemContent += string(toolJSON)
+			if i < len(tools)-1 {
+				firstSystemContent += ", "
+			}
+		}
+		firstSystemContent += "]"
+	}
+
+	// Output first system block if it has content
+	if firstSystemContent != "" {
+		sb.WriteString("<|im_start|>system\n")
+		sb.WriteString(firstSystemContent)
+		sb.WriteString("<|im_end|>\n")
+	}
+
+	// Find the index of the last assistant message for thinking stripping
+	lastAssistantIndex := -1
+	for i := len(messages) - 1; i >= startIdx; i-- {
+		if messages[i].Role == "assistant" {
+			lastAssistantIndex = i
+			break
+		}
+	}
+
+	// Track whether we need to add generation prompt
+	needsGenerationPrompt := len(messages) > 0
+
+	for i := startIdx; i < len(messages); i++ {
+		message := messages[i]
+		switch message.Role {
+		case "system":
+			// Additional system messages (after the first) are rendered normally
+			sb.WriteString("<|im_start|>system\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+
+		case "user":
+			sb.WriteString("<|im_start|>user\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+			needsGenerationPrompt = true
+
+		case "assistant":
+			sb.WriteString("<|im_start|>assistant\n")
+
+			// Check if this is the last assistant message
+			isLastAssistant := i == lastAssistantIndex
+
+			// Process content (may need thinking stripped)
+			content := message.Content
+
+			// Handle thinking tags in assistant content
+			keepPastThinking := r.IsThinking && (thinkValue != nil && thinkValue.Bool())
+			if strings.Contains(content, "</think>") {
+				parts := strings.SplitN(content, "</think>", 2)
+				if len(parts) > 1 {
+					if !isLastAssistant && !keepPastThinking {
+						// Strip thinking entirely for past assistant messages
+						content = strings.TrimSpace(parts[1])
+					} else {
+						// Preserve thinking but trim whitespace after </think>
+						content = parts[0] + "</think>" + strings.TrimLeft(parts[1], " \t\n\r")
+					}
+				}
+			}
+
+			if len(message.ToolCalls) > 0 {
+				// Assistant with tool calls - write content first (if any after stripping)
+				if content != "" {
+					sb.WriteString(content)
+				}
+
+				for _, toolCall := range message.ToolCalls {
+					sb.WriteString("<|tool_call_start|>")
+					toolCallJSON := map[string]any{
+						"name":      toolCall.Function.Name,
+						"arguments": toolCall.Function.Arguments,
+					}
+					callJSON, _ := json.Marshal(toolCallJSON)
+					sb.WriteString(string(callJSON))
+					sb.WriteString("<|tool_call_end|>")
+				}
+			} else {
+				sb.WriteString(content)
+			}
+
+			sb.WriteString("<|im_end|>\n")
+			needsGenerationPrompt = true // Always add gen prompt after assistant when add_generation_prompt=true
+
+		case "tool":
+			// Tool responses are rendered as plain messages per the chat template
+			sb.WriteString("<|im_start|>tool\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+			needsGenerationPrompt = true
+		}
+	}
+
+	// Add generation prompt
+	if needsGenerationPrompt {
+		sb.WriteString("<|im_start|>assistant\n")
+		// Note: Model is a "thinking-only" model - it will output <think> itself
+		// We don't add <think> tag to the prompt
+	}
+
+	return sb.String(), nil
+}

model/renderers/lfm2_test.go

@@ -0,0 +1,427 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestLFM2Renderer(t *testing.T) {
+	tests := []struct {
+		name       string
+		messages   []api.Message
+		tools      []api.Tool
+		thinkValue *api.ThinkValue
+		expected   string
+	}{
+		{
+			name: "basic user message",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "basic with system message",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "multiple system messages rendered separately",
+			messages: []api.Message{
+				{Role: "system", Content: "First instruction."},
+				{Role: "system", Content: "Second instruction."},
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>system\nFirst instruction.<|im_end|>\n<|im_start|>system\nSecond instruction.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "multi-turn conversation",
+			messages: []api.Message{
+				{Role: "user", Content: "What is 2+2?"},
+				{Role: "assistant", Content: "The answer is 4."},
+				{Role: "user", Content: "Thanks!"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nWhat is 2+2?<|im_end|>\n<|im_start|>assistant\nThe answer is 4.<|im_end|>\n<|im_start|>user\nThanks!<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "only system message",
+			messages: []api.Message{
+				{Role: "system", Content: "You are helpful."},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>system\nYou are helpful.<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// When assistant is the LAST assistant, thinking is preserved (even with keep_past_thinking=false)
+			name: "user-assistant-user: last assistant preserves thinking",
+			messages: []api.Message{
+				{Role: "user", Content: "Q1"},
+				{Role: "assistant", Content: "<think>reasoning</think>A1"},
+				{Role: "user", Content: "Q2"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nQ1<|im_end|>\n<|im_start|>assistant\n<think>reasoning</think>A1<|im_end|>\n<|im_start|>user\nQ2<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// With two assistants, first is stripped (not last), second preserved (is last)
+			name: "multi-turn thinking: first stripped, second preserved",
+			messages: []api.Message{
+				{Role: "user", Content: "Q1"},
+				{Role: "assistant", Content: "<think>reason1</think>A1"},
+				{Role: "user", Content: "Q2"},
+				{Role: "assistant", Content: "<think>reason2</think>A2"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nQ1<|im_end|>\n<|im_start|>assistant\nA1<|im_end|>\n<|im_start|>user\nQ2<|im_end|>\n<|im_start|>assistant\n<think>reason2</think>A2<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// With thinking enabled (keep_past_thinking=true), both preserved
+			name: "multi-turn thinking: both preserved when thinking enabled",
+			messages: []api.Message{
+				{Role: "user", Content: "Q1"},
+				{Role: "assistant", Content: "<think>reason1</think>A1"},
+				{Role: "user", Content: "Q2"},
+				{Role: "assistant", Content: "<think>reason2</think>A2"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   "<|im_start|>user\nQ1<|im_end|>\n<|im_start|>assistant\n<think>reason1</think>A1<|im_end|>\n<|im_start|>user\nQ2<|im_end|>\n<|im_start|>assistant\n<think>reason2</think>A2<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "assistant with tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<|im_start|>user` + "\n" + `What's the weather?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n" + `<|tool_call_start|>{"arguments":{"location":"Paris"},"name":"get_weather"}<|tool_call_end|><|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
+		},
+		{
+			name: "assistant with content and tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role:    "assistant",
+					Content: "Let me check.",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<|im_start|>user` + "\n" + `What's the weather in Paris?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n" + `Let me check.<|tool_call_start|>{"arguments":{"location":"Paris"},"name":"get_weather"}<|tool_call_end|><|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
+		},
+		{
+			name: "tool response",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{Role: "assistant", Content: "Let me check."},
+				{Role: "tool", Content: "22C, Sunny"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\nLet me check.<|im_end|>\n<|im_start|>tool\n22C, Sunny<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "multiple tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "Get weather for Paris and London"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "London",
+								}),
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<|im_start|>user` + "\n" + `Get weather for Paris and London<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n" + `<|tool_call_start|>{"arguments":{"location":"Paris"},"name":"get_weather"}<|tool_call_end|><|tool_call_start|>{"arguments":{"location":"London"},"name":"get_weather"}<|tool_call_end|><|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
+		},
+		{
+			name: "tools definitions with system message",
+			messages: []api.Message{
+				{Role: "system", Content: "You are helpful."},
+				{Role: "user", Content: "What's the weather?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: testPropsMap(map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							}),
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<|im_start|>system` + "\n" + `You are helpful.` + "\n" + `List of tools: [{"type":"function","function":{"name":"get_weather","description":"Get current weather","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}}}]<|im_end|>` + "\n" + `<|im_start|>user` + "\n" + `What's the weather?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
+		},
+		{
+			name: "tools definitions without system message",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: testPropsMap(map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							}),
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<|im_start|>system` + "\n" + `List of tools: [{"type":"function","function":{"name":"get_weather","description":"Get current weather","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}}}]<|im_end|>` + "\n" + `<|im_start|>user` + "\n" + `What's the weather?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
+		},
+		{
+			name: "multiple tools without system message",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+					},
+				},
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_time",
+						Description: "Get time",
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>system\nList of tools: [{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get weather\",\"parameters\":{\"type\":\"\",\"properties\":null}}}, {\"type\":\"function\",\"function\":{\"name\":\"get_time\",\"description\":\"Get time\",\"parameters\":{\"type\":\"\",\"properties\":null}}}]<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "user-tool sequence",
+			messages: []api.Message{
+				{Role: "user", Content: "Check weather"},
+				{Role: "tool", Content: "22C"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nCheck weather<|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "full tool call cycle",
+			messages: []api.Message{
+				{Role: "user", Content: "Check weather"},
+				{Role: "assistant", Content: "Let me check"},
+				{Role: "tool", Content: "22C"},
+				{Role: "assistant", Content: "It's 22C"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nCheck weather<|im_end|>\n<|im_start|>assistant\nLet me check<|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\nIt's 22C<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "unicode content",
+			messages: []api.Message{
+				{Role: "user", Content: "你好世界! مرحبا 🌍"},
+				{Role: "assistant", Content: "Hello! 👋"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\n你好世界! مرحبا 🌍<|im_end|>\n<|im_start|>assistant\nHello! 👋<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "newlines in content",
+			messages: []api.Message{
+				{Role: "user", Content: "Line 1\nLine 2\n\nLine 4"},
+				{Role: "assistant", Content: "Response with\nmultiple\nlines"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nLine 1\nLine 2\n\nLine 4<|im_end|>\n<|im_start|>assistant\nResponse with\nmultiple\nlines<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			name: "empty assistant content",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: ""},
+				{Role: "user", Content: "OK"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n<|im_end|>\n<|im_start|>user\nOK<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// Generation prompt does NOT include <think> - model outputs it
+			name: "generation prompt has no think tag",
+			messages: []api.Message{
+				{Role: "user", Content: "Think hard"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   "<|im_start|>user\nThink hard<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// Interleaved: thinking before tool call - last assistant preserves thinking
+			name: "thinking before tool call (last assistant)",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role:    "assistant",
+					Content: "<think>I need to check the weather</think>",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\n<think>I need to check the weather</think><|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// Two assistants with tool calls - first has thinking stripped
+			name: "two assistants with tools: first thinking stripped",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role:    "assistant",
+					Content: "<think>checking</think>",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "22C"},
+				{Role: "assistant", Content: "<think>got result</think>It's 22C!"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\n<|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\n<think>got result</think>It's 22C!<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// Two assistants with tools - both preserved when thinking enabled
+			name: "two assistants with tools: both preserved when thinking enabled",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role:    "assistant",
+					Content: "<think>checking</think>",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "22C"},
+				{Role: "assistant", Content: "<think>got result</think>It's 22C!"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\n<think>checking</think><|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\n<think>got result</think>It's 22C!<|im_end|>\n<|im_start|>assistant\n",
+		},
+		{
+			// Content before thinking before tool call
+			name: "content then thinking then tool call",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role:    "assistant",
+					Content: "Let me check.<think>Using weather API</think>",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: testArgs(map[string]any{
+									"location": "Paris",
+								}),
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\nLet me check.<think>Using weather API</think><|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>assistant\n",
+		},
+	}
+
+	renderer := &LFM2Renderer{IsThinking: true}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
+			if err != nil {
+				t.Fatalf("Render() error = %v", err)
+			}
+			if diff := cmp.Diff(tt.expected, rendered); diff != "" {
+				t.Errorf("Render() mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}

model/renderers/renderer.go

@@ -82,6 +82,10 @@ func rendererForName(name string) Renderer {
 		return &FunctionGemmaRenderer{}
 	case "glm-4.7":
 		return &GLM47Renderer{}
+	case "lfm2":
+		return &LFM2Renderer{IsThinking: false}
+	case "lfm2-thinking":
+		return &LFM2Renderer{IsThinking: true}
 	default:
 		return nil
 	}

openai/openai.go

@@ -794,3 +794,47 @@ func ToImageGenerationResponse(resp api.GenerateResponse) ImageGenerationRespons
 		Data:    data,
 	}
 }
+
+// ImageEditRequest is an OpenAI-compatible image edit request.
+type ImageEditRequest struct {
+	Model  string `json:"model"`
+	Prompt string `json:"prompt"`
+	Image  string `json:"image"`           // Base64-encoded image data
+	Size   string `json:"size,omitempty"`  // e.g., "1024x1024"
+	Seed   *int64 `json:"seed,omitempty"`
+}
+
+// FromImageEditRequest converts an OpenAI image edit request to an Ollama GenerateRequest.
+func FromImageEditRequest(r ImageEditRequest) (api.GenerateRequest, error) {
+	req := api.GenerateRequest{
+		Model:  r.Model,
+		Prompt: r.Prompt,
+	}
+
+	// Decode the input image
+	if r.Image != "" {
+		imgData, err := decodeImageURL(r.Image)
+		if err != nil {
+			return api.GenerateRequest{}, fmt.Errorf("invalid image: %w", err)
+		}
+		req.Images = append(req.Images, imgData)
+	}
+
+	// Parse size if provided (e.g., "1024x768")
+	if r.Size != "" {
+		var w, h int32
+		if _, err := fmt.Sscanf(r.Size, "%dx%d", &w, &h); err == nil {
+			req.Width = w
+			req.Height = h
+		}
+	}
+
+	if r.Seed != nil {
+		if req.Options == nil {
+			req.Options = map[string]any{}
+		}
+		req.Options["seed"] = *r.Seed
+	}
+
+	return req, nil
+}

openai/openai_test.go

@@ -448,3 +448,86 @@ func TestFromChatRequest_TopLogprobsRange(t *testing.T) {
 		})
 	}
 }
+
+func TestFromImageEditRequest_Basic(t *testing.T) {
+	req := ImageEditRequest{
+		Model:  "test-model",
+		Prompt: "make it blue",
+		Image:  prefix + image,
+	}
+
+	result, err := FromImageEditRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Model != "test-model" {
+		t.Errorf("expected model 'test-model', got %q", result.Model)
+	}
+
+	if result.Prompt != "make it blue" {
+		t.Errorf("expected prompt 'make it blue', got %q", result.Prompt)
+	}
+
+	if len(result.Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(result.Images))
+	}
+}
+
+func TestFromImageEditRequest_WithSize(t *testing.T) {
+	req := ImageEditRequest{
+		Model:  "test-model",
+		Prompt: "make it blue",
+		Image:  prefix + image,
+		Size:   "512x768",
+	}
+
+	result, err := FromImageEditRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Width != 512 {
+		t.Errorf("expected width 512, got %d", result.Width)
+	}
+
+	if result.Height != 768 {
+		t.Errorf("expected height 768, got %d", result.Height)
+	}
+}
+
+func TestFromImageEditRequest_WithSeed(t *testing.T) {
+	seed := int64(12345)
+	req := ImageEditRequest{
+		Model:  "test-model",
+		Prompt: "make it blue",
+		Image:  prefix + image,
+		Seed:   &seed,
+	}
+
+	result, err := FromImageEditRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Options == nil {
+		t.Fatal("expected options to be set")
+	}
+
+	if result.Options["seed"] != seed {
+		t.Errorf("expected seed %d, got %v", seed, result.Options["seed"])
+	}
+}
+
+func TestFromImageEditRequest_InvalidImage(t *testing.T) {
+	req := ImageEditRequest{
+		Model:  "test-model",
+		Prompt: "make it blue",
+		Image:  "not-valid-base64",
+	}
+
+	_, err := FromImageEditRequest(req)
+	if err == nil {
+		t.Error("expected error for invalid image")
+	}
+}

server/create.go

@@ -28,6 +28,7 @@ import (
 	"github.com/ollama/ollama/format"
 	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@@ -90,7 +91,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 			ch <- resp
 		}
 
-		oldManifest, _ := ParseNamedManifest(name)
+		oldManifest, _ := manifest.ParseNamedManifest(name)
 
 		var baseLayers []*layerGGML
 		var err error
@@ -123,9 +124,9 @@ func (s *Server) CreateHandler(c *gin.Context) {
 				}
 
 				if err == nil && !remote && (config.Renderer == "" || config.Parser == "" || config.Requires == "") {
-					manifest, mErr := ParseNamedManifest(fromName)
-					if mErr == nil && manifest.Config.Digest != "" {
-						configPath, pErr := GetBlobsPath(manifest.Config.Digest)
+					mf, mErr := manifest.ParseNamedManifest(fromName)
+					if mErr == nil && mf.Config.Digest != "" {
+						configPath, pErr := manifest.BlobsPath(mf.Config.Digest)
 						if pErr == nil {
 							if cfgFile, fErr := os.Open(configPath); fErr == nil {
 								var baseConfig model.ConfigV2
@@ -342,7 +343,7 @@ func detectModelTypeFromFiles(files map[string]string) string {
 			return "gguf"
 		} else {
 			// try to see if we can find a gguf file even without the file extension
-			blobPath, err := GetBlobsPath(files[fn])
+			blobPath, err := manifest.BlobsPath(files[fn])
 			if err != nil {
 				slog.Error("error getting blobs path", "file", fn)
 				return ""
@@ -394,7 +395,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 			return nil, fmt.Errorf("%w: %s: %s", errFilePath, err, fp)
 		}
 
-		blobPath, err := GetBlobsPath(digest)
+		blobPath, err := manifest.BlobsPath(digest)
 		if err != nil {
 			return nil, err
 		}
@@ -432,7 +433,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 		return nil, err
 	}
 
-	layer, err := NewLayer(t, mediaType)
+	layer, err := manifest.NewLayer(t, mediaType)
 	if err != nil {
 		return nil, err
 	}
@@ -465,7 +466,7 @@ func kvFromLayers(baseLayers []*layerGGML) (ofs.Config, error) {
 }
 
 func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML, config *model.ConfigV2, fn func(resp api.ProgressResponse)) (err error) {
-	var layers []Layer
+	var layers []manifest.Layer
 	for _, layer := range baseLayers {
 		if layer.GGML != nil {
 			quantType := strings.ToUpper(cmp.Or(r.Quantize, r.Quantization))
@@ -550,13 +551,13 @@ func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML,
 	}
 
 	for _, layer := range layers {
-		if layer.status != "" {
-			fn(api.ProgressResponse{Status: layer.status})
+		if layer.Status != "" {
+			fn(api.ProgressResponse{Status: layer.Status})
 		}
 	}
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	if err := WriteManifest(name, *configLayer, layers); err != nil {
+	if err := manifest.WriteManifest(name, *configLayer, layers); err != nil {
 		return err
 	}
 
@@ -577,7 +578,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 		return nil, err
 	}
 
-	blob, err := GetBlobsPath(layer.Digest)
+	blob, err := manifest.BlobsPath(layer.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -599,7 +600,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 	}
 	temp.Seek(0, io.SeekStart)
 	fn(api.ProgressResponse{Status: "verifying conversion"})
-	newLayer, err := NewLayer(temp, layer.MediaType)
+	newLayer, err := manifest.NewLayer(temp, layer.MediaType)
 	if err != nil {
 		return nil, err
 	}
@@ -619,7 +620,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 	var layers []*layerGGML
 
 	fn(api.ProgressResponse{Status: "parsing GGUF"})
-	blobPath, err := GetBlobsPath(digest)
+	blobPath, err := manifest.BlobsPath(digest)
 	if err != nil {
 		return nil, err
 	}
@@ -654,7 +655,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 		mediatype = "application/vnd.ollama.image.projector"
 	}
 
-	layer, err := NewLayerFromLayer(digest, mediatype, blob.Name())
+	layer, err := manifest.NewLayerFromLayer(digest, mediatype, blob.Name())
 	if err != nil {
 		slog.Debug("could not create new layer from layer", "error", err)
 		return nil, err
@@ -665,8 +666,8 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 	return detectChatTemplate(layers)
 }
 
-func removeLayer(layers []Layer, mediatype string) []Layer {
-	return slices.DeleteFunc(layers, func(layer Layer) bool {
+func removeLayer(layers []manifest.Layer, mediatype string) []manifest.Layer {
+	return slices.DeleteFunc(layers, func(layer manifest.Layer) bool {
 		if layer.MediaType != mediatype {
 			return false
 		}
@@ -680,7 +681,7 @@ func removeLayer(layers []Layer, mediatype string) []Layer {
 	})
 }
 
-func setTemplate(layers []Layer, t string) ([]Layer, error) {
+func setTemplate(layers []manifest.Layer, t string) ([]manifest.Layer, error) {
 	layers = removeLayer(layers, "application/vnd.ollama.image.template")
 	if _, err := template.Parse(t); err != nil {
 		return nil, fmt.Errorf("%w: %s", errBadTemplate, err)
@@ -690,7 +691,7 @@ func setTemplate(layers []Layer, t string) ([]Layer, error) {
 	}
 
 	blob := strings.NewReader(t)
-	layer, err := NewLayer(blob, "application/vnd.ollama.image.template")
+	layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.template")
 	if err != nil {
 		return nil, err
 	}
@@ -699,11 +700,11 @@ func setTemplate(layers []Layer, t string) ([]Layer, error) {
 	return layers, nil
 }
 
-func setSystem(layers []Layer, s string) ([]Layer, error) {
+func setSystem(layers []manifest.Layer, s string) ([]manifest.Layer, error) {
 	layers = removeLayer(layers, "application/vnd.ollama.image.system")
 	if s != "" {
 		blob := strings.NewReader(s)
-		layer, err := NewLayer(blob, "application/vnd.ollama.image.system")
+		layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.system")
 		if err != nil {
 			return nil, err
 		}
@@ -712,9 +713,9 @@ func setSystem(layers []Layer, s string) ([]Layer, error) {
 	return layers, nil
 }
 
-func setLicense(layers []Layer, l string) ([]Layer, error) {
+func setLicense(layers []manifest.Layer, l string) ([]manifest.Layer, error) {
 	blob := strings.NewReader(l)
-	layer, err := NewLayer(blob, "application/vnd.ollama.image.license")
+	layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.license")
 	if err != nil {
 		return nil, err
 	}
@@ -722,7 +723,7 @@ func setLicense(layers []Layer, l string) ([]Layer, error) {
 	return layers, nil
 }
 
-func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
+func setParameters(layers []manifest.Layer, p map[string]any) ([]manifest.Layer, error) {
 	if p == nil {
 		p = make(map[string]any)
 	}
@@ -731,7 +732,7 @@ func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 			continue
 		}
 
-		digestPath, err := GetBlobsPath(layer.Digest)
+		digestPath, err := manifest.BlobsPath(layer.Digest)
 		if err != nil {
 			return nil, err
 		}
@@ -765,7 +766,7 @@ func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 	if err := json.NewEncoder(&b).Encode(p); err != nil {
 		return nil, err
 	}
-	layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
+	layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.params")
 	if err != nil {
 		return nil, err
 	}
@@ -773,7 +774,7 @@ func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 	return layers, nil
 }
 
-func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
+func setMessages(layers []manifest.Layer, m []api.Message) ([]manifest.Layer, error) {
 	// this leaves the old messages intact if no new messages were specified
 	// which may not be the correct behaviour
 	if len(m) == 0 {
@@ -786,7 +787,7 @@ func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
 	if err := json.NewEncoder(&b).Encode(m); err != nil {
 		return nil, err
 	}
-	layer, err := NewLayer(&b, "application/vnd.ollama.image.messages")
+	layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.messages")
 	if err != nil {
 		return nil, err
 	}
@@ -794,7 +795,7 @@ func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
 	return layers, nil
 }
 
-func createConfigLayer(layers []Layer, config model.ConfigV2) (*Layer, error) {
+func createConfigLayer(layers []manifest.Layer, config model.ConfigV2) (*manifest.Layer, error) {
 	digests := make([]string, len(layers))
 	for i, layer := range layers {
 		digests[i] = layer.Digest
@@ -805,7 +806,7 @@ func createConfigLayer(layers []Layer, config model.ConfigV2) (*Layer, error) {
 	if err := json.NewEncoder(&b).Encode(config); err != nil {
 		return nil, err
 	}
-	layer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	layer, err := manifest.NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		return nil, err
 	}

server/create_test.go

@@ -10,6 +10,7 @@ import (
 	"testing"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/manifest"
 )
 
 func TestConvertFromSafetensors(t *testing.T) {
@@ -17,7 +18,7 @@ func TestConvertFromSafetensors(t *testing.T) {
 
 	// Helper function to create a new layer and return its digest
 	makeTemp := func(content string) string {
-		l, err := NewLayer(strings.NewReader(content), "application/octet-stream")
+		l, err := manifest.NewLayer(strings.NewReader(content), "application/octet-stream")
 		if err != nil {
 			t.Fatalf("Failed to create layer: %v", err)
 		}

server/download.go

@@ -24,6 +24,8 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
 )
 
 const maxRetries = 6
@@ -456,7 +458,7 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 }
 
 type downloadOpts struct {
-	mp      ModelPath
+	n       model.Name
 	digest  string
 	regOpts *registryOptions
 	fn      func(api.ProgressResponse)
@@ -465,10 +467,10 @@ type downloadOpts struct {
 // downloadBlob downloads a blob from the registry and stores it in the blobs directory
 func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ error) {
 	if opts.digest == "" {
-		return false, fmt.Errorf(("%s: %s"), opts.mp.GetNamespaceRepository(), "digest is empty")
+		return false, fmt.Errorf(("%s: %s"), opts.n.DisplayNamespaceModel(), "digest is empty")
 	}
 
-	fp, err := GetBlobsPath(opts.digest)
+	fp, err := manifest.BlobsPath(opts.digest)
 	if err != nil {
 		return false, err
 	}
@@ -492,8 +494,8 @@ func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ erro
 	data, ok := blobDownloadManager.LoadOrStore(opts.digest, &blobDownload{Name: fp, Digest: opts.digest})
 	download := data.(*blobDownload)
 	if !ok {
-		requestURL := opts.mp.BaseURL()
-		requestURL = requestURL.JoinPath("v2", opts.mp.GetNamespaceRepository(), "blobs", opts.digest)
+		requestURL := opts.n.BaseURL()
+		requestURL = requestURL.JoinPath("v2", opts.n.DisplayNamespaceModel(), "blobs", opts.digest)
 		if err := download.Prepare(ctx, requestURL, opts.regOpts); err != nil {
 			blobDownloadManager.Delete(opts.digest)
 			return false, err

server/images.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"context"
 	"crypto/sha256"
-	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -24,6 +23,7 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/gguf"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/template"
@@ -274,44 +274,22 @@ func (m *Model) String() string {
 	return modelfile.String()
 }
 
-func GetManifest(mp ModelPath) (*Manifest, string, error) {
-	fp, err := mp.GetManifestPath()
-	if err != nil {
-		return nil, "", err
-	}
-
-	f, err := os.Open(fp)
-	if err != nil {
-		return nil, "", err
-	}
-	defer f.Close()
-
-	sha256sum := sha256.New()
-
-	var manifest Manifest
-	if err := json.NewDecoder(io.TeeReader(f, sha256sum)).Decode(&manifest); err != nil {
-		return nil, "", err
-	}
-
-	return &manifest, hex.EncodeToString(sha256sum.Sum(nil)), nil
-}
-
 func GetModel(name string) (*Model, error) {
-	mp := ParseModelPath(name)
-	manifest, digest, err := GetManifest(mp)
+	n := model.ParseName(name)
+	mf, err := manifest.ParseNamedManifest(n)
 	if err != nil {
 		return nil, err
 	}
 
-	model := &Model{
-		Name:      mp.GetFullTagname(),
-		ShortName: mp.GetShortTagname(),
-		Digest:    digest,
+	m := &Model{
+		Name:      n.String(),
+		ShortName: n.DisplayShortest(),
+		Digest:    mf.Digest(),
 		Template:  template.DefaultTemplate,
 	}
 
-	if manifest.Config.Digest != "" {
-		filename, err := GetBlobsPath(manifest.Config.Digest)
+	if mf.Config.Digest != "" {
+		filename, err := manifest.BlobsPath(mf.Config.Digest)
 		if err != nil {
 			return nil, err
 		}
@@ -322,29 +300,29 @@ func GetModel(name string) (*Model, error) {
 		}
 		defer configFile.Close()
 
-		if err := json.NewDecoder(configFile).Decode(&model.Config); err != nil {
+		if err := json.NewDecoder(configFile).Decode(&m.Config); err != nil {
 			return nil, err
 		}
 	}
 
-	for _, layer := range manifest.Layers {
-		filename, err := GetBlobsPath(layer.Digest)
+	for _, layer := range mf.Layers {
+		filename, err := manifest.BlobsPath(layer.Digest)
 		if err != nil {
 			return nil, err
 		}
 
 		switch layer.MediaType {
 		case "application/vnd.ollama.image.model":
-			model.ModelPath = filename
-			model.ParentModel = layer.From
+			m.ModelPath = filename
+			m.ParentModel = layer.From
 		case "application/vnd.ollama.image.embed":
 			// Deprecated in versions  > 0.1.2
 			// TODO: remove this warning in a future version
 			slog.Info("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
 		case "application/vnd.ollama.image.adapter":
-			model.AdapterPaths = append(model.AdapterPaths, filename)
+			m.AdapterPaths = append(m.AdapterPaths, filename)
 		case "application/vnd.ollama.image.projector":
-			model.ProjectorPaths = append(model.ProjectorPaths, filename)
+			m.ProjectorPaths = append(m.ProjectorPaths, filename)
 		case "application/vnd.ollama.image.prompt",
 			"application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
@@ -352,7 +330,7 @@ func GetModel(name string) (*Model, error) {
 				return nil, err
 			}
 
-			model.Template, err = template.Parse(string(bts))
+			m.Template, err = template.Parse(string(bts))
 			if err != nil {
 				return nil, err
 			}
@@ -362,7 +340,7 @@ func GetModel(name string) (*Model, error) {
 				return nil, err
 			}
 
-			model.System = string(bts)
+			m.System = string(bts)
 		case "application/vnd.ollama.image.params":
 			params, err := os.Open(filename)
 			if err != nil {
@@ -371,7 +349,7 @@ func GetModel(name string) (*Model, error) {
 			defer params.Close()
 
 			// parse model options parameters into a map so that we can see which fields have been specified explicitly
-			if err = json.NewDecoder(params).Decode(&model.Options); err != nil {
+			if err = json.NewDecoder(params).Decode(&m.Options); err != nil {
 				return nil, err
 			}
 		case "application/vnd.ollama.image.messages":
@@ -381,7 +359,7 @@ func GetModel(name string) (*Model, error) {
 			}
 			defer msgs.Close()
 
-			if err = json.NewDecoder(msgs).Decode(&model.Messages); err != nil {
+			if err = json.NewDecoder(msgs).Decode(&m.Messages); err != nil {
 				return nil, err
 			}
 		case "application/vnd.ollama.image.license":
@@ -389,11 +367,11 @@ func GetModel(name string) (*Model, error) {
 			if err != nil {
 				return nil, err
 			}
-			model.License = append(model.License, string(bts))
+			m.License = append(m.License, string(bts))
 		}
 	}
 
-	return model, nil
+	return m, nil
 }
 
 func CopyModel(src, dst model.Name) error {
@@ -408,7 +386,7 @@ func CopyModel(src, dst model.Name) error {
 		return nil
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := manifest.Path()
 	if err != nil {
 		return err
 	}
@@ -437,7 +415,7 @@ func CopyModel(src, dst model.Name) error {
 
 func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 	// Ignore corrupt manifests to avoid blocking deletion of layers that are freshly orphaned
-	manifests, err := Manifests(true)
+	manifests, err := manifest.Manifests(true)
 	if err != nil {
 		return err
 	}
@@ -452,7 +430,7 @@ func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 
 	// only delete the files which are still in the deleteMap
 	for k := range deleteMap {
-		fp, err := GetBlobsPath(k)
+		fp, err := manifest.BlobsPath(k)
 		if err != nil {
 			slog.Info(fmt.Sprintf("couldn't get file path for '%s': %v", k, err))
 			continue
@@ -468,7 +446,7 @@ func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 
 func PruneLayers() error {
 	deleteMap := make(map[string]struct{})
-	p, err := GetBlobsPath("")
+	p, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
@@ -483,9 +461,9 @@ func PruneLayers() error {
 		name := blob.Name()
 		name = strings.ReplaceAll(name, "-", ":")
 
-		_, err := GetBlobsPath(name)
+		_, err := manifest.BlobsPath(name)
 		if err != nil {
-			if errors.Is(err, ErrInvalidDigestFormat) {
+			if errors.Is(err, manifest.ErrInvalidDigestFormat) {
 				// remove invalid blobs (e.g. partial downloads)
 				if err := os.Remove(filepath.Join(p, blob.Name())); err != nil {
 					slog.Error("couldn't remove blob", "blob", blob.Name(), "error", err)
@@ -510,63 +488,30 @@ func PruneLayers() error {
 	return nil
 }
 
-func PruneDirectory(path string) error {
-	info, err := os.Lstat(path)
-	if err != nil {
-		return err
-	}
-
-	if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
-		entries, err := os.ReadDir(path)
-		if err != nil {
-			return err
-		}
-
-		for _, entry := range entries {
-			if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
-				return err
-			}
-		}
-
-		entries, err = os.ReadDir(path)
-		if err != nil {
-			return err
-		}
-
-		if len(entries) > 0 {
-			return nil
-		}
-
-		return os.Remove(path)
-	}
-
-	return nil
-}
-
 func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
-	mp := ParseModelPath(name)
+	n := model.ParseName(name)
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
 
-	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
+	if n.ProtocolScheme == "http" && !regOpts.Insecure {
 		return errInsecureProtocol
 	}
 
-	manifest, _, err := GetManifest(mp)
+	mf, err := manifest.ParseNamedManifest(n)
 	if err != nil {
 		fn(api.ProgressResponse{Status: "couldn't retrieve manifest"})
 		return err
 	}
 
-	var layers []Layer
-	layers = append(layers, manifest.Layers...)
-	if manifest.Config.Digest != "" {
-		layers = append(layers, manifest.Config)
+	var layers []manifest.Layer
+	layers = append(layers, mf.Layers...)
+	if mf.Config.Digest != "" {
+		layers = append(layers, mf.Config)
 	}
 
 	// Use fast transfer for models with tensor layers (many small blobs)
 	if hasTensorLayers(layers) {
 		// Read raw manifest JSON to preserve tensor metadata fields
-		manifestPath, err := mp.GetManifestPath()
+		manifestPath, err := manifest.PathForName(n)
 		if err != nil {
 			return err
 		}
@@ -574,7 +519,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		if err != nil {
 			return err
 		}
-		if err := pushWithTransfer(ctx, mp, layers, manifestJSON, regOpts, fn); err != nil {
+		if err := pushWithTransfer(ctx, n, layers, manifestJSON, regOpts, fn); err != nil {
 			return err
 		}
 		fn(api.ProgressResponse{Status: "success"})
@@ -582,17 +527,17 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	}
 
 	for _, layer := range layers {
-		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
+		if err := uploadBlob(ctx, n, layer, regOpts, fn); err != nil {
 			slog.Info(fmt.Sprintf("error uploading blob: %v", err))
 			return err
 		}
 	}
 
 	fn(api.ProgressResponse{Status: "pushing manifest"})
-	requestURL := mp.BaseURL()
-	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
+	requestURL := n.BaseURL()
+	requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "manifests", n.Tag)
 
-	manifestJSON, err := json.Marshal(manifest)
+	manifestJSON, err := json.Marshal(mf)
 	if err != nil {
 		return err
 	}
@@ -611,44 +556,44 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 }
 
 func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
-	mp := ParseModelPath(name)
+	n := model.ParseName(name)
 
 	// build deleteMap to prune unused layers
 	deleteMap := make(map[string]struct{})
-	manifest, _, err := GetManifest(mp)
+	existingMf, err := manifest.ParseNamedManifest(n)
 	if errors.Is(err, os.ErrNotExist) {
 		// noop
 	} else if err != nil {
 		slog.Warn("pulling model with bad existing manifest", "name", name, "error", err)
 	} else {
-		for _, l := range manifest.Layers {
+		for _, l := range existingMf.Layers {
 			deleteMap[l.Digest] = struct{}{}
 		}
-		if manifest.Config.Digest != "" {
-			deleteMap[manifest.Config.Digest] = struct{}{}
+		if existingMf.Config.Digest != "" {
+			deleteMap[existingMf.Config.Digest] = struct{}{}
 		}
 	}
 
-	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
+	if n.ProtocolScheme == "http" && !regOpts.Insecure {
 		return errInsecureProtocol
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})
 
-	manifest, err = pullModelManifest(ctx, mp, regOpts)
+	mf, err := pullModelManifest(ctx, n, regOpts)
 	if err != nil {
 		return fmt.Errorf("pull model manifest: %s", err)
 	}
 
-	var layers []Layer
-	layers = append(layers, manifest.Layers...)
-	if manifest.Config.Digest != "" {
-		layers = append(layers, manifest.Config)
+	var layers []manifest.Layer
+	layers = append(layers, mf.Layers...)
+	if mf.Config.Digest != "" {
+		layers = append(layers, mf.Config)
 	}
 
 	// Use fast transfer for models with tensor layers (many small blobs)
 	if hasTensorLayers(layers) {
-		if err := pullWithTransfer(ctx, mp, layers, manifest, regOpts, fn); err != nil {
+		if err := pullWithTransfer(ctx, n, layers, mf, regOpts, fn); err != nil {
 			return err
 		}
 		fn(api.ProgressResponse{Status: "success"})
@@ -658,7 +603,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	skipVerify := make(map[string]bool)
 	for _, layer := range layers {
 		cacheHit, err := downloadBlob(ctx, downloadOpts{
-			mp:      mp,
+			n:       n,
 			digest:  layer.Digest,
 			regOpts: regOpts,
 			fn:      fn,
@@ -677,7 +622,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		}
 		if err := verifyBlob(layer.Digest); err != nil {
 			if errors.Is(err, errDigestMismatch) {
-				fp, err := GetBlobsPath(layer.Digest)
+				fp, err := manifest.BlobsPath(layer.Digest)
 				if err != nil {
 					return err
 				}
@@ -692,16 +637,16 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	for _, layer := range layers {
 		delete(deleteMap, layer.Digest)
 	}
-	delete(deleteMap, manifest.Config.Digest)
+	delete(deleteMap, mf.Config.Digest)
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
 
-	manifestJSON, err := json.Marshal(manifest)
+	manifestJSON, err := json.Marshal(mf)
 	if err != nil {
 		return err
 	}
 
-	fp, err := mp.GetManifestPath()
+	fp, err := manifest.PathForName(n)
 	if err != nil {
 		return err
 	}
@@ -728,9 +673,9 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 }
 
 // hasTensorLayers checks if any layer has tensor media type.
-func hasTensorLayers(layers []Layer) bool {
+func hasTensorLayers(layers []manifest.Layer) bool {
 	for _, layer := range layers {
-		if layer.MediaType == MediaTypeImageTensor {
+		if layer.MediaType == manifest.MediaTypeImageTensor {
 			return true
 		}
 	}
@@ -738,7 +683,7 @@ func hasTensorLayers(layers []Layer) bool {
 }
 
 // pullWithTransfer uses the simplified x/transfer package for downloading blobs.
-func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifest *Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer, mf *manifest.Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	blobs := make([]transfer.Blob, len(layers))
 	for i, layer := range layers {
 		blobs[i] = transfer.Blob{
@@ -747,12 +692,12 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		}
 	}
 
-	destDir, err := GetBlobsPath("")
+	destDir, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
 
-	base := mp.BaseURL()
+	base := n.BaseURL()
 	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		base.Scheme = "http"
 	}
@@ -784,7 +729,7 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		Blobs:      blobs,
 		BaseURL:    baseURL,
 		DestDir:    destDir,
-		Repository: mp.GetNamespaceRepository(),
+		Repository: n.DisplayNamespaceModel(),
 		Progress:   progress,
 		Token:      regOpts.Token,
 		GetToken:   getToken,
@@ -795,12 +740,12 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 
 	// Write manifest
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	manifestJSON, err := json.Marshal(manifest)
+	manifestJSON, err := json.Marshal(mf)
 	if err != nil {
 		return err
 	}
 
-	fp, err := mp.GetManifestPath()
+	fp, err := manifest.PathForName(n)
 	if err != nil {
 		return err
 	}
@@ -812,7 +757,7 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 }
 
 // pushWithTransfer uses the simplified x/transfer package for uploading blobs and manifest.
-func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+func pushWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	blobs := make([]transfer.Blob, len(layers))
 	for i, layer := range layers {
 		blobs[i] = transfer.Blob{
@@ -822,12 +767,12 @@ func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		}
 	}
 
-	srcDir, err := GetBlobsPath("")
+	srcDir, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
 
-	base := mp.BaseURL()
+	base := n.BaseURL()
 	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		base.Scheme = "http"
 	}
@@ -864,13 +809,13 @@ func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		GetToken:    getToken,
 		Logger:      slog.Default(),
 		Manifest:    manifestJSON,
-		ManifestRef: mp.Tag,
-		Repository:  mp.GetNamespaceRepository(),
+		ManifestRef: n.Tag,
+		Repository:  n.DisplayNamespaceModel(),
 	})
 }
 
-func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
-	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
+func pullModelManifest(ctx context.Context, n model.Name, regOpts *registryOptions) (*manifest.Manifest, error) {
+	requestURL := n.BaseURL().JoinPath("v2", n.DisplayNamespaceModel(), "manifests", n.Tag)
 
 	headers := make(http.Header)
 	headers.Set("Accept", "application/vnd.docker.distribution.manifest.v2+json")
@@ -880,7 +825,7 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio
 	}
 	defer resp.Body.Close()
 
-	var m Manifest
+	var m manifest.Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}
@@ -1042,7 +987,7 @@ func parseRegistryChallenge(authStr string) registryChallenge {
 var errDigestMismatch = errors.New("digest mismatch, file must be downloaded again")
 
 func verifyBlob(digest string) error {
-	fp, err := GetBlobsPath(digest)
+	fp, err := manifest.BlobsPath(digest)
 	if err != nil {
 		return err
 	}

server/model.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/model"
 )
@@ -20,19 +21,19 @@ import (
 var intermediateBlobs map[string]string = make(map[string]string)
 
 type layerGGML struct {
-	Layer
+	manifest.Layer
 	*ggml.GGML
 }
 
 func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
-	m, err := ParseNamedManifest(name)
+	m, err := manifest.ParseNamedManifest(name)
 	switch {
 	case errors.Is(err, os.ErrNotExist):
 		if err := PullModel(ctx, name.String(), &registryOptions{}, fn); err != nil {
 			return nil, err
 		}
 
-		m, err = ParseNamedManifest(name)
+		m, err = manifest.ParseNamedManifest(name)
 		if err != nil {
 			return nil, err
 		}
@@ -41,7 +42,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	}
 
 	for _, layer := range m.Layers {
-		layer, err := NewLayerFromLayer(layer.Digest, layer.MediaType, name.DisplayShortest())
+		layer, err := manifest.NewLayerFromLayer(layer.Digest, layer.MediaType, name.DisplayShortest())
 		if err != nil {
 			return nil, err
 		}
@@ -50,7 +51,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 		case "application/vnd.ollama.image.model",
 			"application/vnd.ollama.image.projector",
 			"application/vnd.ollama.image.adapter":
-			blobpath, err := GetBlobsPath(layer.Digest)
+			blobpath, err := manifest.BlobsPath(layer.Digest)
 			if err != nil {
 				return nil, err
 			}
@@ -81,12 +82,12 @@ func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 			if t, err := template.Named(s); err != nil {
 				slog.Debug("template detection", "error", err, "template", s)
 			} else {
-				layer, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
+				layer, err := manifest.NewLayer(t.Reader(), "application/vnd.ollama.image.template")
 				if err != nil {
 					return nil, err
 				}
 
-				layer.status = fmt.Sprintf("using autodetected template %s", t.Name)
+				layer.Status = fmt.Sprintf("using autodetected template %s", t.Name)
 				layers = append(layers, &layerGGML{layer, nil})
 
 				if t.Parameters != nil {
@@ -95,7 +96,7 @@ func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 						return nil, err
 					}
 
-					layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
+					layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.params")
 					if err != nil {
 						return nil, err
 					}

server/modelpath.go

@@ -1,146 +0,0 @@
-package server
-
-import (
-	"errors"
-	"fmt"
-	"io/fs"
-	"net/url"
-	"os"
-	"path/filepath"
-	"regexp"
-	"strings"
-
-	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/types/model"
-)
-
-type ModelPath struct {
-	ProtocolScheme string
-	Registry       string
-	Namespace      string
-	Repository     string
-	Tag            string
-}
-
-const (
-	DefaultRegistry       = "registry.ollama.ai"
-	DefaultNamespace      = "library"
-	DefaultTag            = "latest"
-	DefaultProtocolScheme = "https"
-)
-
-var (
-	ErrInvalidImageFormat  = errors.New("invalid image format")
-	ErrInvalidDigestFormat = errors.New("invalid digest format")
-	ErrInvalidProtocol     = errors.New("invalid protocol scheme")
-	ErrInsecureProtocol    = errors.New("insecure protocol http")
-	ErrModelPathInvalid    = errors.New("invalid model path")
-)
-
-func ParseModelPath(name string) ModelPath {
-	mp := ModelPath{
-		ProtocolScheme: DefaultProtocolScheme,
-		Registry:       DefaultRegistry,
-		Namespace:      DefaultNamespace,
-		Repository:     "",
-		Tag:            DefaultTag,
-	}
-
-	before, after, found := strings.Cut(name, "://")
-	if found {
-		mp.ProtocolScheme = before
-		name = after
-	}
-
-	name = strings.ReplaceAll(name, string(os.PathSeparator), "/")
-	parts := strings.Split(name, "/")
-	switch len(parts) {
-	case 3:
-		mp.Registry = parts[0]
-		mp.Namespace = parts[1]
-		mp.Repository = parts[2]
-	case 2:
-		mp.Namespace = parts[0]
-		mp.Repository = parts[1]
-	case 1:
-		mp.Repository = parts[0]
-	}
-
-	if repo, tag, found := strings.Cut(mp.Repository, ":"); found {
-		mp.Repository = repo
-		mp.Tag = tag
-	}
-
-	return mp
-}
-
-func (mp ModelPath) GetNamespaceRepository() string {
-	return fmt.Sprintf("%s/%s", mp.Namespace, mp.Repository)
-}
-
-func (mp ModelPath) GetFullTagname() string {
-	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
-}
-
-func (mp ModelPath) GetShortTagname() string {
-	if mp.Registry == DefaultRegistry {
-		if mp.Namespace == DefaultNamespace {
-			return fmt.Sprintf("%s:%s", mp.Repository, mp.Tag)
-		}
-		return fmt.Sprintf("%s/%s:%s", mp.Namespace, mp.Repository, mp.Tag)
-	}
-	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
-}
-
-// GetManifestPath returns the path to the manifest file for the given model path, it is up to the caller to create the directory if it does not exist.
-func (mp ModelPath) GetManifestPath() (string, error) {
-	name := model.Name{
-		Host:      mp.Registry,
-		Namespace: mp.Namespace,
-		Model:     mp.Repository,
-		Tag:       mp.Tag,
-	}
-	if !name.IsValid() {
-		return "", fs.ErrNotExist
-	}
-	return filepath.Join(envconfig.Models(), "manifests", name.Filepath()), nil
-}
-
-func (mp ModelPath) BaseURL() *url.URL {
-	return &url.URL{
-		Scheme: mp.ProtocolScheme,
-		Host:   mp.Registry,
-	}
-}
-
-func GetManifestPath() (string, error) {
-	path := filepath.Join(envconfig.Models(), "manifests")
-	if err := os.MkdirAll(path, 0o755); err != nil {
-		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
-	}
-
-	return path, nil
-}
-
-func GetBlobsPath(digest string) (string, error) {
-	// only accept actual sha256 digests
-	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
-	re := regexp.MustCompile(pattern)
-
-	if digest != "" && !re.MatchString(digest) {
-		return "", ErrInvalidDigestFormat
-	}
-
-	digest = strings.ReplaceAll(digest, ":", "-")
-	path := filepath.Join(envconfig.Models(), "blobs", digest)
-	dirPath := filepath.Dir(path)
-	if digest == "" {
-		dirPath = path
-	}
-
-	if err := os.MkdirAll(dirPath, 0o755); err != nil {
-		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
-	}
-
-	return path, nil
-}

server/modelpath_test.go

@@ -1,153 +0,0 @@
-package server
-
-import (
-	"path/filepath"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestGetBlobsPath(t *testing.T) {
-	// GetBlobsPath expects an actual directory to exist
-	tempDir := t.TempDir()
-
-	tests := []struct {
-		name     string
-		digest   string
-		expected string
-		err      error
-	}{
-		{
-			"empty digest",
-			"",
-			filepath.Join(tempDir, "blobs"),
-			nil,
-		},
-		{
-			"valid with colon",
-			"sha256:456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9",
-			filepath.Join(tempDir, "blobs", "sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9"),
-			nil,
-		},
-		{
-			"valid with dash",
-			"sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9",
-			filepath.Join(tempDir, "blobs", "sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9"),
-			nil,
-		},
-		{
-			"digest too short",
-			"sha256-45640291",
-			"",
-			ErrInvalidDigestFormat,
-		},
-		{
-			"digest too long",
-			"sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9aaaaaaaaaa",
-			"",
-			ErrInvalidDigestFormat,
-		},
-		{
-			"digest invalid chars",
-			"../sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7a",
-			"",
-			ErrInvalidDigestFormat,
-		},
-	}
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			t.Setenv("OLLAMA_MODELS", tempDir)
-
-			got, err := GetBlobsPath(tc.digest)
-
-			require.ErrorIs(t, tc.err, err, tc.name)
-			assert.Equal(t, tc.expected, got, tc.name)
-		})
-	}
-}
-
-func TestParseModelPath(t *testing.T) {
-	tests := []struct {
-		name string
-		arg  string
-		want ModelPath
-	}{
-		{
-			"full path https",
-			"https://example.com/ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       "example.com",
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"full path http",
-			"http://example.com/ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "http",
-				Registry:       "example.com",
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no protocol",
-			"example.com/ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       "example.com",
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no registry",
-			"ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       DefaultRegistry,
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no namespace",
-			"repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       DefaultRegistry,
-				Namespace:      DefaultNamespace,
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no tag",
-			"repo",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       DefaultRegistry,
-				Namespace:      DefaultNamespace,
-				Repository:     "repo",
-				Tag:            DefaultTag,
-			},
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			got := ParseModelPath(tc.arg)
-
-			if got != tc.want {
-				t.Errorf("got: %q want: %q", got, tc.want)
-			}
-		})
-	}
-}

server/quantization.go

@@ -198,8 +198,8 @@ func newType(t *fsggml.Tensor, kv fsggml.KV, qs *quantizeState, ftype fsggml.Fil
 	name := t.Name
 	quantize := strings.HasSuffix(name, "weight")
 
-	// don't quantize vision stuff
-	quantize = quantize && (!strings.Contains(name, "v.") || strings.Contains(name, "_v."))
+	// don't quantize vision encoder tensors (named with "v." prefix)
+	quantize = quantize && !strings.HasPrefix(name, "v.")
 	quantize = quantize && !strings.Contains(name, "mm.")
 
 	// quantize only 2D and 3D tensors (experts)
@@ -219,6 +219,9 @@ func newType(t *fsggml.Tensor, kv fsggml.KV, qs *quantizeState, ftype fsggml.Fil
 	// NOTE: can't use LLM_TN here because the layer number is not known
 	quantize = quantize && !strings.Contains(name, "ssm_conv1d.weight")
 
+	// do not quantize LFM2's shortconv kernel weights
+	quantize = quantize && !strings.Contains(name, "shortconv.conv.weight")
+
 	// do not quantize RWKV's time_mix_first tensors
 	quantize = quantize && !strings.Contains(name, "time_mix_first.weight")
 	quantize = quantize && !strings.Contains(name, "time_mix_w1.weight")

server/routes.go

@@ -39,6 +39,7 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/middleware"
 	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/model/renderers"
@@ -974,7 +975,7 @@ func (s *Server) PushHandler(c *gin.Context) {
 // is.
 func getExistingName(n model.Name) (model.Name, error) {
 	var zero model.Name
-	existing, err := Manifests(true)
+	existing, err := manifest.Manifests(true)
 	if err != nil {
 		return zero, err
 	}
@@ -1018,7 +1019,7 @@ func (s *Server) DeleteHandler(c *gin.Context) {
 		return
 	}
 
-	m, err := ParseNamedManifest(n)
+	m, err := manifest.ParseNamedManifest(n)
 	if err != nil {
 		switch {
 		case os.IsNotExist(err):
@@ -1080,7 +1081,7 @@ func (s *Server) ShowHandler(c *gin.Context) {
 func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
-		return nil, ErrModelPathInvalid
+		return nil, model.Unqualified(name)
 	}
 	name, err := getExistingName(name)
 	if err != nil {
@@ -1112,7 +1113,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 
 	// For safetensors LLM models (experimental), populate details from config.json
 	if m.Config.ModelFormat == "safetensors" && slices.Contains(m.Config.Capabilities, "completion") {
-		if info, err := xserver.GetSafetensorsLLMInfo(name.String()); err == nil {
+		if info, err := xserver.GetSafetensorsLLMInfo(name); err == nil {
 			if arch, ok := info["general.architecture"].(string); ok && arch != "" {
 				modelDetails.Family = arch
 			}
@@ -1121,7 +1122,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 			}
 		}
 		// Get torch_dtype directly from config.json for quantization level
-		if dtype, err := xserver.GetSafetensorsDtype(name.String()); err == nil && dtype != "" {
+		if dtype, err := xserver.GetSafetensorsDtype(name); err == nil && dtype != "" {
 			modelDetails.QuantizationLevel = dtype
 		}
 	}
@@ -1135,7 +1136,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		msgs[i] = api.Message{Role: msg.Role, Content: msg.Content}
 	}
 
-	manifest, err := ParseNamedManifest(name)
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return nil, err
 	}
@@ -1147,8 +1148,11 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		Details:      modelDetails,
 		Messages:     msgs,
 		Capabilities: m.Capabilities(),
-		ModifiedAt:   manifest.fi.ModTime(),
+		ModifiedAt:   mf.FileInfo().ModTime(),
 		Requires:     m.Config.Requires,
+		// Several integrations crash on a nil/omitempty+empty ModelInfo, so by
+		// default we return an empty map.
+		ModelInfo: make(map[string]any),
 	}
 
 	if m.Config.RemoteHost != "" {
@@ -1211,7 +1215,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	if slices.Contains(m.Capabilities(), model.CapabilityImage) {
 		// Populate tensor info if verbose
 		if req.Verbose {
-			if tensors, err := xserver.GetSafetensorsTensorInfo(name.String()); err == nil {
+			if tensors, err := xserver.GetSafetensorsTensorInfo(name); err == nil {
 				resp.Tensors = tensors
 			}
 		}
@@ -1220,12 +1224,12 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 
 	// For safetensors LLM models (experimental), populate ModelInfo from config.json
 	if m.Config.ModelFormat == "safetensors" && slices.Contains(m.Config.Capabilities, "completion") {
-		if info, err := xserver.GetSafetensorsLLMInfo(name.String()); err == nil {
+		if info, err := xserver.GetSafetensorsLLMInfo(name); err == nil {
 			resp.ModelInfo = info
 		}
 		// Populate tensor info if verbose
 		if req.Verbose {
-			if tensors, err := xserver.GetSafetensorsTensorInfo(name.String()); err == nil {
+			if tensors, err := xserver.GetSafetensorsTensorInfo(name); err == nil {
 				resp.Tensors = tensors
 			}
 		}
@@ -1282,7 +1286,7 @@ func getModelData(digest string, verbose bool) (ggml.KV, ggml.Tensors, error) {
 }
 
 func (s *Server) ListHandler(c *gin.Context) {
-	ms, err := Manifests(true)
+	ms, err := manifest.Manifests(true)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -1313,8 +1317,8 @@ func (s *Server) ListHandler(c *gin.Context) {
 			RemoteModel: cf.RemoteModel,
 			RemoteHost:  cf.RemoteHost,
 			Size:        m.Size(),
-			Digest:      m.digest,
-			ModifiedAt:  m.fi.ModTime(),
+			Digest:      m.Digest(),
+			ModifiedAt:  m.FileInfo().ModTime(),
 			Details: api.ModelDetails{
 				Format:            cf.ModelFormat,
 				Family:            cf.ModelFamily,
@@ -1373,7 +1377,7 @@ func (s *Server) CopyHandler(c *gin.Context) {
 }
 
 func (s *Server) HeadBlobHandler(c *gin.Context) {
-	path, err := GetBlobsPath(c.Param("digest"))
+	path, err := manifest.BlobsPath(c.Param("digest"))
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -1389,7 +1393,7 @@ func (s *Server) HeadBlobHandler(c *gin.Context) {
 
 func (s *Server) CreateBlobHandler(c *gin.Context) {
 	if ib, ok := intermediateBlobs[c.Param("digest")]; ok {
-		p, err := GetBlobsPath(ib)
+		p, err := manifest.BlobsPath(ib)
 		if err != nil {
 			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 			return
@@ -1407,7 +1411,7 @@ func (s *Server) CreateBlobHandler(c *gin.Context) {
 		}
 	}
 
-	path, err := GetBlobsPath(c.Param("digest"))
+	path, err := manifest.BlobsPath(c.Param("digest"))
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -1425,7 +1429,7 @@ func (s *Server) CreateBlobHandler(c *gin.Context) {
 		return
 	}
 
-	layer, err := NewLayer(c.Request.Body, "")
+	layer, err := manifest.NewLayer(c.Request.Body, "")
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -1600,8 +1604,9 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
-	// OpenAI-compatible image generation endpoint
+	// OpenAI-compatible image generation endpoints
 	r.POST("/v1/images/generations", middleware.ImageGenerationsMiddleware(), s.GenerateHandler)
+	r.POST("/v1/images/edits", middleware.ImageEditsMiddleware(), s.GenerateHandler)
 
 	// Inference (Anthropic compatibility)
 	r.POST("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)
@@ -1625,7 +1630,7 @@ func Serve(ln net.Listener) error {
 	slog.SetDefault(logutil.NewLogger(os.Stderr, envconfig.LogLevel()))
 	slog.Info("server config", "env", envconfig.Values())
 
-	blobsDir, err := GetBlobsPath("")
+	blobsDir, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
@@ -1634,7 +1639,7 @@ func Serve(ln net.Listener) error {
 	}
 
 	if !envconfig.NoPrune() {
-		if _, err := Manifests(false); err != nil {
+		if _, err := manifest.Manifests(false); err != nil {
 			slog.Warn("corrupt manifests detected, skipping prune operation.  Re-pull or delete to clear", "error", err)
 		} else {
 			// clean up unused layers and manifests
@@ -1642,12 +1647,12 @@ func Serve(ln net.Listener) error {
 				return err
 			}
 
-			manifestsPath, err := GetManifestPath()
+			manifestsPath, err := manifest.Path()
 			if err != nil {
 				return err
 			}
 
-			if err := PruneDirectory(manifestsPath); err != nil {
+			if err := manifest.PruneDirectory(manifestsPath); err != nil {
 				return err
 			}
 		}
@@ -2519,6 +2524,11 @@ func (s *Server) handleImageGenerate(c *gin.Context, req api.GenerateRequest, mo
 		}
 	}
 
+	var images []llm.ImageData
+	for i, imgData := range req.Images {
+		images = append(images, llm.ImageData{ID: i, Data: imgData})
+	}
+
 	var streamStarted bool
 	if err := runner.Completion(c.Request.Context(), llm.CompletionRequest{
 		Prompt: req.Prompt,
@@ -2526,6 +2536,7 @@ func (s *Server) handleImageGenerate(c *gin.Context, req api.GenerateRequest, mo
 		Height: req.Height,
 		Steps:  req.Steps,
 		Seed:   seed,
+		Images: images,
 	}, func(cr llm.CompletionResponse) {
 		streamStarted = true
 		res := api.GenerateResponse{

server/routes_create_test.go

@@ -25,6 +25,7 @@ import (
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -223,15 +224,15 @@ func TestCreateFromModelInheritsRendererParser(t *testing.T) {
 		t.Fatalf("expected status code 200, actual %d", w.Code)
 	}
 
-	manifest, err := ParseNamedManifest(model.ParseName("child"))
+	mf, err := manifest.ParseNamedManifest(model.ParseName("child"))
 	if err != nil {
 		t.Fatalf("parse manifest: %v", err)
 	}
-	if manifest.Config.Digest == "" {
+	if mf.Config.Digest == "" {
 		t.Fatalf("unexpected empty config digest for child manifest")
 	}
 
-	configPath, err := GetBlobsPath(manifest.Config.Digest)
+	configPath, err := manifest.BlobsPath(mf.Config.Digest)
 	if err != nil {
 		t.Fatalf("config blob path: %v", err)
 	}

server/routes_delete_test.go

@@ -10,6 +10,7 @@ import (
 	"github.com/gin-gonic/gin"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -93,13 +94,13 @@ func TestDeleteDuplicateLayers(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	config, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	config, err := manifest.NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		t.Fatal(err)
 	}
 
 	// create a manifest with duplicate layers
-	if err := WriteManifest(n, config, []Layer{config}); err != nil {
+	if err := manifest.WriteManifest(n, config, []manifest.Layer{config}); err != nil {
 		t.Fatal(err)
 	}
 

server/routes_generate_test.go

@@ -2193,3 +2193,157 @@ func TestGenerateUnload(t *testing.T) {
 		}
 	})
 }
+
+func TestGenerateWithImages(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	mock := mockRunner{
+		CompletionResponse: llm.CompletionResponse{
+			Done:               true,
+			DoneReason:         llm.DoneReasonStop,
+			PromptEvalCount:    1,
+			PromptEvalDuration: 1,
+			EvalCount:          1,
+			EvalDuration:       1,
+		},
+	}
+
+	s := Server{
+		sched: &Scheduler{
+			pendingReqCh:    make(chan *LlmRequest, 1),
+			finishedReqCh:   make(chan *LlmRequest, 1),
+			expiredCh:       make(chan *runnerRef, 1),
+			unloadedCh:      make(chan any, 1),
+			loaded:          make(map[string]*runnerRef),
+			newServerFn:     newMockServer(&mock),
+			getGpuFn:        getGpuFn,
+			getSystemInfoFn: getSystemInfoFn,
+			waitForRecovery: 250 * time.Millisecond,
+			loadFn: func(req *LlmRequest, _ *ggml.GGML, _ ml.SystemInfo, _ []ml.DeviceInfo, _ bool) bool {
+				time.Sleep(time.Millisecond)
+				req.successCh <- &runnerRef{
+					llama: &mock,
+				}
+				return false
+			},
+		},
+	}
+
+	go s.sched.Run(t.Context())
+
+	_, digest := createBinFile(t, ggml.KV{
+		"general.architecture":          "llama",
+		"llama.block_count":             uint32(1),
+		"llama.context_length":          uint32(8192),
+		"llama.embedding_length":        uint32(4096),
+		"llama.attention.head_count":    uint32(32),
+		"llama.attention.head_count_kv": uint32(8),
+		"tokenizer.ggml.tokens":         []string{""},
+		"tokenizer.ggml.scores":         []float32{0},
+		"tokenizer.ggml.token_type":     []int32{0},
+	}, []*ggml.Tensor{
+		{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+	})
+
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model:  "test",
+		Files:  map[string]string{"file.gguf": digest},
+		Stream: &stream,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("images passed to completion request", func(t *testing.T) {
+		testImage := []byte("test-image-data")
+
+		mock.CompletionResponse.Content = "Image processed"
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "Describe this image",
+			Images: []api.ImageData{testImage},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+
+		// Verify images were passed to the completion request
+		if len(mock.CompletionRequest.Images) != 1 {
+			t.Fatalf("expected 1 image in completion request, got %d", len(mock.CompletionRequest.Images))
+		}
+
+		if !bytes.Equal(mock.CompletionRequest.Images[0].Data, testImage) {
+			t.Errorf("image data mismatch in completion request")
+		}
+
+		if mock.CompletionRequest.Images[0].ID != 0 {
+			t.Errorf("expected image ID 0, got %d", mock.CompletionRequest.Images[0].ID)
+		}
+	})
+
+	t.Run("multiple images passed to completion request", func(t *testing.T) {
+		testImage1 := []byte("test-image-1")
+		testImage2 := []byte("test-image-2")
+
+		mock.CompletionResponse.Content = "Images processed"
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "Compare these images",
+			Images: []api.ImageData{testImage1, testImage2},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+
+		// Verify both images were passed
+		if len(mock.CompletionRequest.Images) != 2 {
+			t.Fatalf("expected 2 images in completion request, got %d", len(mock.CompletionRequest.Images))
+		}
+
+		if !bytes.Equal(mock.CompletionRequest.Images[0].Data, testImage1) {
+			t.Errorf("first image data mismatch")
+		}
+
+		if !bytes.Equal(mock.CompletionRequest.Images[1].Data, testImage2) {
+			t.Errorf("second image data mismatch")
+		}
+
+		if mock.CompletionRequest.Images[0].ID != 0 || mock.CompletionRequest.Images[1].ID != 1 {
+			t.Errorf("expected image IDs 0 and 1, got %d and %d",
+				mock.CompletionRequest.Images[0].ID, mock.CompletionRequest.Images[1].ID)
+		}
+	})
+
+	t.Run("no images when none provided", func(t *testing.T) {
+		mock.CompletionResponse.Content = "No images"
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "Hello",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+
+		// Verify no images in completion request
+		if len(mock.CompletionRequest.Images) != 0 {
+			t.Fatalf("expected 0 images in completion request, got %d", len(mock.CompletionRequest.Images))
+		}
+	})
+}

server/upload.go

@@ -21,12 +21,14 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
 )
 
 var blobUploadManager sync.Map
 
 type blobUpload struct {
-	Layer
+	manifest.Layer
 
 	Total     int64
 	Completed atomic.Int64
@@ -51,7 +53,7 @@ const (
 )
 
 func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *registryOptions) error {
-	p, err := GetBlobsPath(b.Digest)
+	p, err := manifest.BlobsPath(b.Digest)
 	if err != nil {
 		return err
 	}
@@ -59,7 +61,7 @@ func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *reg
 	if b.From != "" {
 		values := requestURL.Query()
 		values.Add("mount", b.Digest)
-		values.Add("from", ParseModelPath(b.From).GetNamespaceRepository())
+		values.Add("from", model.ParseName(b.From).DisplayNamespaceModel())
 		requestURL.RawQuery = values.Encode()
 	}
 
@@ -128,7 +130,7 @@ func (b *blobUpload) Run(ctx context.Context, opts *registryOptions) {
 	defer blobUploadManager.Delete(b.Digest)
 	ctx, b.CancelFunc = context.WithCancel(ctx)
 
-	p, err := GetBlobsPath(b.Digest)
+	p, err := manifest.BlobsPath(b.Digest)
 	if err != nil {
 		b.err = err
 		return
@@ -364,9 +366,9 @@ func (p *progressWriter) Rollback() {
 	p.written = 0
 }
 
-func uploadBlob(ctx context.Context, mp ModelPath, layer Layer, opts *registryOptions, fn func(api.ProgressResponse)) error {
-	requestURL := mp.BaseURL()
-	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs", layer.Digest)
+func uploadBlob(ctx context.Context, n model.Name, layer manifest.Layer, opts *registryOptions, fn func(api.ProgressResponse)) error {
+	requestURL := n.BaseURL()
+	requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "blobs", layer.Digest)
 
 	resp, err := makeRequestWithRetry(ctx, http.MethodHead, requestURL, nil, nil, opts)
 	switch {
@@ -388,8 +390,8 @@ func uploadBlob(ctx context.Context, mp ModelPath, layer Layer, opts *registryOp
 	data, ok := blobUploadManager.LoadOrStore(layer.Digest, &blobUpload{Layer: layer})
 	upload := data.(*blobUpload)
 	if !ok {
-		requestURL := mp.BaseURL()
-		requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs/uploads/")
+		requestURL := n.BaseURL()
+		requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "blobs/uploads/")
 		if err := upload.Prepare(ctx, requestURL, opts); err != nil {
 			blobUploadManager.Delete(layer.Digest)
 			return err

types/model/name.go

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"net/url"
 	"path/filepath"
 	"strings"
 )
@@ -35,22 +36,25 @@ func Unqualified(n Name) error {
 const MissingPart = "!MISSING!"
 
 const (
-	defaultHost      = "registry.ollama.ai"
-	defaultNamespace = "library"
-	defaultTag       = "latest"
+	defaultHost           = "registry.ollama.ai"
+	defaultNamespace      = "library"
+	defaultTag            = "latest"
+	defaultProtocolScheme = "https"
 )
 
 // DefaultName returns a name with the default values for the host, namespace,
-// and tag parts. The model and digest parts are empty.
+// tag, and protocol scheme parts. The model and digest parts are empty.
 //
 //   - The default host is ("registry.ollama.ai")
 //   - The default namespace is ("library")
 //   - The default tag is ("latest")
+//   - The default protocol scheme is ("https")
 func DefaultName() Name {
 	return Name{
-		Host:      defaultHost,
-		Namespace: defaultNamespace,
-		Tag:       defaultTag,
+		Host:           defaultHost,
+		Namespace:      defaultNamespace,
+		Tag:            defaultTag,
+		ProtocolScheme: defaultProtocolScheme,
 	}
 }
 
@@ -87,10 +91,11 @@ func (k partKind) String() string {
 // It is not guaranteed to be valid. Use [Name.IsValid] to check if the name
 // is valid.
 type Name struct {
-	Host      string
-	Namespace string
-	Model     string
-	Tag       string
+	Host           string
+	Namespace      string
+	Model          string
+	Tag            string
+	ProtocolScheme string
 }
 
 // ParseName parses and assembles a Name from a name string. The
@@ -160,7 +165,9 @@ func ParseNameBare(s string) Name {
 	}
 
 	scheme, host, ok := strings.Cut(s, "://")
-	if !ok {
+	if ok {
+		n.ProtocolScheme = scheme
+	} else {
 		host = scheme
 	}
 	n.Host = host
@@ -189,12 +196,13 @@ func ParseNameFromFilepath(s string) (n Name) {
 	return n
 }
 
-// Merge merges the host, namespace, and tag parts of the two names,
+// Merge merges the host, namespace, tag, and protocol scheme parts of the two names,
 // preferring the non-empty parts of a.
 func Merge(a, b Name) Name {
 	a.Host = cmp.Or(a.Host, b.Host)
 	a.Namespace = cmp.Or(a.Namespace, b.Namespace)
 	a.Tag = cmp.Or(a.Tag, b.Tag)
+	a.ProtocolScheme = cmp.Or(a.ProtocolScheme, b.ProtocolScheme)
 	return a
 }
 
@@ -305,6 +313,23 @@ func (n Name) EqualFold(o Name) bool {
 		strings.EqualFold(n.Tag, o.Tag)
 }
 
+// BaseURL returns the base URL for the registry.
+func (n Name) BaseURL() *url.URL {
+	return &url.URL{
+		Scheme: n.ProtocolScheme,
+		Host:   n.Host,
+	}
+}
+
+// DisplayNamespaceModel returns the namespace and model joined by "/".
+func (n Name) DisplayNamespaceModel() string {
+	var b strings.Builder
+	b.WriteString(n.Namespace)
+	b.WriteByte('/')
+	b.WriteString(n.Model)
+	return b.String()
+}
+
 func isValidLen(kind partKind, s string) bool {
 	switch kind {
 	case kindHost:

types/model/name_test.go

@@ -32,10 +32,11 @@ func TestParseNameParts(t *testing.T) {
 		{
 			in: "scheme://host:port/namespace/model:tag",
 			want: Name{
-				Host:      "host:port",
-				Namespace: "namespace",
-				Model:     "model",
-				Tag:       "tag",
+				Host:           "host:port",
+				Namespace:      "namespace",
+				Model:          "model",
+				Tag:            "tag",
+				ProtocolScheme: "scheme",
 			},
 			wantFilepath: filepath.Join("host:port", "namespace", "model", "tag"),
 		},

x/create/client/create.go

@@ -12,9 +12,8 @@ import (
 	"fmt"
 	"io"
 
-	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/progress"
-	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/x/create"
 )
@@ -104,7 +103,7 @@ func CreateModel(opts CreateOptions, p *progress.Progress) error {
 // newLayerCreator returns a LayerCreator callback for creating config/JSON layers.
 func newLayerCreator() create.LayerCreator {
 	return func(r io.Reader, mediaType, name string) (create.LayerInfo, error) {
-		layer, err := server.NewLayer(r, mediaType)
+		layer, err := manifest.NewLayer(r, mediaType)
 		if err != nil {
 			return create.LayerInfo{}, err
 		}
@@ -142,13 +141,13 @@ func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quant
 	}
 
 	// Create layer for quantized weight
-	weightLayer, err := server.NewLayer(bytes.NewReader(qweightData), server.MediaTypeImageTensor)
+	weightLayer, err := manifest.NewLayer(bytes.NewReader(qweightData), manifest.MediaTypeImageTensor)
 	if err != nil {
 		return nil, err
 	}
 
 	// Create layer for scales
-	scalesLayer, err := server.NewLayer(bytes.NewReader(scalesData), server.MediaTypeImageTensor)
+	scalesLayer, err := manifest.NewLayer(bytes.NewReader(scalesData), manifest.MediaTypeImageTensor)
 	if err != nil {
 		return nil, err
 	}
@@ -170,7 +169,7 @@ func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quant
 
 	// Add qbiases layer if present (affine mode)
 	if qbiasData != nil {
-		qbiasLayer, err := server.NewLayer(bytes.NewReader(qbiasData), server.MediaTypeImageTensor)
+		qbiasLayer, err := manifest.NewLayer(bytes.NewReader(qbiasData), manifest.MediaTypeImageTensor)
 		if err != nil {
 			return nil, err
 		}
@@ -187,7 +186,7 @@ func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quant
 
 // createUnquantizedLayer creates a single tensor layer without quantization.
 func createUnquantizedLayer(r io.Reader, name string) ([]create.LayerInfo, error) {
-	layer, err := server.NewLayer(r, server.MediaTypeImageTensor)
+	layer, err := manifest.NewLayer(r, manifest.MediaTypeImageTensor)
 	if err != nil {
 		return nil, err
 	}
@@ -222,15 +221,15 @@ func newManifestWriter(opts CreateOptions, capabilities []string) create.Manifes
 		}
 
 		// Create config layer blob
-		configLayer, err := server.NewLayer(bytes.NewReader(configJSON), "application/vnd.docker.container.image.v1+json")
+		configLayer, err := manifest.NewLayer(bytes.NewReader(configJSON), "application/vnd.docker.container.image.v1+json")
 		if err != nil {
 			return fmt.Errorf("failed to create config layer: %w", err)
 		}
 
-		// Convert LayerInfo to server.Layer
-		serverLayers := make([]server.Layer, 0, len(layers))
+		// Convert LayerInfo to manifest.Layer
+		manifestLayers := make([]manifest.Layer, 0, len(layers))
 		for _, l := range layers {
-			serverLayers = append(serverLayers, server.Layer{
+			manifestLayers = append(manifestLayers, manifest.Layer{
 				MediaType: l.MediaType,
 				Digest:    l.Digest,
 				Size:      l.Size,
@@ -244,19 +243,19 @@ func newManifestWriter(opts CreateOptions, capabilities []string) create.Manifes
 			if err != nil {
 				return err
 			}
-			serverLayers = append(serverLayers, modelfileLayers...)
+			manifestLayers = append(manifestLayers, modelfileLayers...)
 		}
 
-		return server.WriteManifest(name, configLayer, serverLayers)
+		return manifest.WriteManifest(name, configLayer, manifestLayers)
 	}
 }
 
 // createModelfileLayers creates layers for template, system, and license from Modelfile config.
-func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
-	var layers []server.Layer
+func createModelfileLayers(mf *ModelfileConfig) ([]manifest.Layer, error) {
+	var layers []manifest.Layer
 
 	if mf.Template != "" {
-		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.Template)), "application/vnd.ollama.image.template")
+		layer, err := manifest.NewLayer(bytes.NewReader([]byte(mf.Template)), "application/vnd.ollama.image.template")
 		if err != nil {
 			return nil, fmt.Errorf("failed to create template layer: %w", err)
 		}
@@ -264,7 +263,7 @@ func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
 	}
 
 	if mf.System != "" {
-		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.System)), "application/vnd.ollama.image.system")
+		layer, err := manifest.NewLayer(bytes.NewReader([]byte(mf.System)), "application/vnd.ollama.image.system")
 		if err != nil {
 			return nil, fmt.Errorf("failed to create system layer: %w", err)
 		}
@@ -272,7 +271,7 @@ func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
 	}
 
 	if mf.License != "" {
-		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.License)), "application/vnd.ollama.image.license")
+		layer, err := manifest.NewLayer(bytes.NewReader([]byte(mf.License)), "application/vnd.ollama.image.license")
 		if err != nil {
 			return nil, fmt.Errorf("failed to create license layer: %w", err)
 		}
@@ -281,19 +280,3 @@ func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
 
 	return layers, nil
 }
-
-// ExtractModelfileConfig extracts template, system, and license from a parsed Modelfile.
-func ExtractModelfileConfig(modelfile *parser.Modelfile) *ModelfileConfig {
-	mfConfig := &ModelfileConfig{}
-	for _, cmd := range modelfile.Commands {
-		switch cmd.Name {
-		case "template":
-			mfConfig.Template = cmd.Args
-		case "system":
-			mfConfig.System = cmd.Args
-		case "license":
-			mfConfig.License = cmd.Args
-		}
-	}
-	return mfConfig
-}

x/imagegen/manifest.go

@@ -6,8 +6,9 @@ import (
 	"io"
 	"os"
 	"path/filepath"
-	"runtime"
 	"strings"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 // ManifestLayer represents a layer in the manifest.
@@ -32,31 +33,15 @@ type ModelManifest struct {
 	BlobDir  string
 }
 
-// DefaultBlobDir returns the default blob storage directory.
 func DefaultBlobDir() string {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		home = "."
-	}
-	switch runtime.GOOS {
-	case "darwin":
-		return filepath.Join(home, ".ollama", "models", "blobs")
-	case "linux":
-		return filepath.Join(home, ".ollama", "models", "blobs")
-	case "windows":
-		return filepath.Join(home, ".ollama", "models", "blobs")
-	default:
-		return filepath.Join(home, ".ollama", "models", "blobs")
-	}
+	return filepath.Join(envconfig.Models(), "blobs")
 }
 
-// DefaultManifestDir returns the default manifest storage directory.
+// DefaultManifestDir returns the manifest storage directory.
+// Respects OLLAMA_MODELS.
+
 func DefaultManifestDir() string {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		home = "."
-	}
-	return filepath.Join(home, ".ollama", "models", "manifests")
+	return filepath.Join(envconfig.Models(), "manifests")
 }
 
 // LoadManifest loads a manifest for the given model name.

x/imagegen/manifest_test.go

@@ -0,0 +1,26 @@
+package imagegen
+
+import (
+	"path/filepath"
+	"testing"
+)
+
+func TestManifestAndBlobDirsRespectOLLAMAModels(t *testing.T) {
+	modelsDir := filepath.Join(t.TempDir(), "models")
+
+	// Simulate packaged/systemd environment
+	t.Setenv("OLLAMA_MODELS", modelsDir)
+	t.Setenv("HOME", "/usr/share/ollama")
+
+	// Manifest dir must respect OLLAMA_MODELS
+	wantManifest := filepath.Join(modelsDir, "manifests")
+	if got := DefaultManifestDir(); got != wantManifest {
+		t.Fatalf("DefaultManifestDir() = %q, want %q", got, wantManifest)
+	}
+
+	// Blob dir must respect OLLAMA_MODELS
+	wantBlobs := filepath.Join(modelsDir, "blobs")
+	if got := DefaultBlobDir(); got != wantBlobs {
+		t.Fatalf("DefaultBlobDir() = %q, want %q", got, wantBlobs)
+	}
+}

x/imagegen/models/flux2/flux2.go

@@ -177,6 +177,20 @@ func (m *Model) GenerateImage(ctx context.Context, prompt string, width, height
 	})
 }
 
+// GenerateImageWithInputs implements runner.ImageEditModel interface.
+// It generates an image conditioned on the provided input images for image editing.
+func (m *Model) GenerateImageWithInputs(ctx context.Context, prompt string, width, height int32, steps int, seed int64, inputImages []image.Image, progress func(step, total int)) (*mlx.Array, error) {
+	return m.GenerateFromConfig(ctx, &GenerateConfig{
+		Prompt:      prompt,
+		Width:       width,
+		Height:      height,
+		Steps:       steps,
+		Seed:        seed,
+		InputImages: inputImages,
+		Progress:    progress,
+	})
+}
+
 // MaxOutputPixels is the maximum output resolution (4 megapixels, ~2048x2048)
 const MaxOutputPixels = 2048 * 2048
 

x/imagegen/runner/runner.go

@@ -9,6 +9,7 @@ import (
 	"encoding/json"
 	"flag"
 	"fmt"
+	"image"
 	"log/slog"
 	"net/http"
 	"os"
@@ -25,11 +26,12 @@ import (
 
 // Request is the image generation request format
 type Request struct {
-	Prompt string `json:"prompt"`
-	Width  int32  `json:"width,omitempty"`
-	Height int32  `json:"height,omitempty"`
-	Steps  int    `json:"steps,omitempty"`
-	Seed   int64  `json:"seed,omitempty"`
+	Prompt string   `json:"prompt"`
+	Width  int32    `json:"width,omitempty"`
+	Height int32    `json:"height,omitempty"`
+	Steps  int      `json:"steps,omitempty"`
+	Seed   int64    `json:"seed,omitempty"`
+	Images [][]byte `json:"images,omitempty"` // Input images for image editing/conditioning
 }
 
 // Response is streamed back for each progress update
@@ -46,6 +48,13 @@ type ImageModel interface {
 	GenerateImage(ctx context.Context, prompt string, width, height int32, steps int, seed int64, progress func(step, total int)) (*mlx.Array, error)
 }
 
+// ImageEditModel extends ImageModel with image editing/conditioning capability.
+// Models that support input images for editing should implement this interface.
+type ImageEditModel interface {
+	ImageModel
+	GenerateImageWithInputs(ctx context.Context, prompt string, width, height int32, steps int, seed int64, inputImages []image.Image, progress func(step, total int)) (*mlx.Array, error)
+}
+
 // Server holds the model and handles requests
 type Server struct {
 	mu        sync.Mutex
@@ -161,6 +170,29 @@ func (s *Server) completionHandler(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// Validate and decode input images
+	const maxInputImages = 2
+	if len(req.Images) > maxInputImages {
+		http.Error(w, fmt.Sprintf("too many input images, maximum is %d", maxInputImages), http.StatusBadRequest)
+		return
+	}
+
+	var inputImages []image.Image
+	if len(req.Images) > 0 {
+		// TODO: add memory check for input images
+
+		inputImages = make([]image.Image, len(req.Images))
+		for i, imgBytes := range req.Images {
+			img, err := imagegen.DecodeImage(imgBytes)
+			if err != nil {
+				http.Error(w, fmt.Sprintf("invalid image %d: %v", i, err), http.StatusBadRequest)
+				return
+			}
+			inputImages[i] = img
+		}
+		slog.Info("decoded input images", "count", len(inputImages))
+	}
+
 	// Serialize generation requests - MLX model may not handle concurrent generation
 	s.mu.Lock()
 	defer s.mu.Unlock()
@@ -192,7 +224,19 @@ func (s *Server) completionHandler(w http.ResponseWriter, r *http.Request) {
 		flusher.Flush()
 	}
 
-	img, err := s.model.GenerateImage(ctx, req.Prompt, req.Width, req.Height, req.Steps, req.Seed, progress)
+	// Use ImageEditModel if available and images provided, otherwise use basic ImageModel
+	var img *mlx.Array
+	var err error
+	if len(inputImages) > 0 {
+		editModel, ok := s.model.(ImageEditModel)
+		if !ok {
+			http.Error(w, "model does not support image editing", http.StatusBadRequest)
+			return
+		}
+		img, err = editModel.GenerateImageWithInputs(ctx, req.Prompt, req.Width, req.Height, req.Steps, req.Seed, inputImages, progress)
+	} else {
+		img, err = s.model.GenerateImage(ctx, req.Prompt, req.Width, req.Height, req.Steps, req.Seed, progress)
+	}
 
 	if err != nil {
 		// Don't send error for cancellation

x/imagegen/server.go

@@ -226,19 +226,27 @@ func (s *Server) Completion(ctx context.Context, req llm.CompletionRequest, fn f
 		seed = time.Now().UnixNano()
 	}
 
+	// Extract raw image bytes from llm.ImageData slice
+	var images [][]byte
+	for _, img := range req.Images {
+		images = append(images, img.Data)
+	}
+
 	// Build request for subprocess
 	creq := struct {
-		Prompt string `json:"prompt"`
-		Width  int32  `json:"width,omitempty"`
-		Height int32  `json:"height,omitempty"`
-		Steps  int32  `json:"steps,omitempty"`
-		Seed   int64  `json:"seed,omitempty"`
+		Prompt string   `json:"prompt"`
+		Width  int32    `json:"width,omitempty"`
+		Height int32    `json:"height,omitempty"`
+		Steps  int32    `json:"steps,omitempty"`
+		Seed   int64    `json:"seed,omitempty"`
+		Images [][]byte `json:"images,omitempty"`
 	}{
 		Prompt: req.Prompt,
 		Width:  req.Width,
 		Height: req.Height,
 		Steps:  req.Steps,
 		Seed:   seed,
+		Images: images,
 	}
 
 	body, err := json.Marshal(creq)

x/server/show.go

@@ -9,7 +9,8 @@ import (
 	"strings"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
 )
 
 // modelConfig represents the HuggingFace config.json structure
@@ -35,22 +36,22 @@ type modelConfig struct {
 
 // GetSafetensorsLLMInfo extracts model information from safetensors LLM models.
 // It reads the config.json layer and returns a map compatible with GGML's KV format.
-func GetSafetensorsLLMInfo(modelName string) (map[string]any, error) {
-	manifest, err := imagegen.LoadManifest(modelName)
+func GetSafetensorsLLMInfo(name model.Name) (map[string]any, error) {
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return nil, fmt.Errorf("failed to load manifest: %w", err)
 	}
 
 	var config modelConfig
-	if err := manifest.ReadConfigJSON("config.json", &config); err != nil {
+	if err := mf.ReadConfigJSON("config.json", &config); err != nil {
 		return nil, fmt.Errorf("failed to read config.json: %w", err)
 	}
 
 	// Calculate total tensor bytes from manifest layers
 	var totalBytes int64
 	var tensorCount int64
-	for _, layer := range manifest.Manifest.Layers {
-		if layer.MediaType == "application/vnd.ollama.image.tensor" {
+	for _, layer := range mf.Layers {
+		if layer.MediaType == manifest.MediaTypeImageTensor {
 			totalBytes += layer.Size
 			tensorCount++
 		}
@@ -151,27 +152,30 @@ func buildModelInfo(config modelConfig, totalTensorBytes, tensorCount int64) map
 
 // GetSafetensorsTensorInfo extracts tensor information from safetensors model layers.
 // Each tensor is stored as a minimal safetensors file with an 88-byte header containing metadata.
-func GetSafetensorsTensorInfo(modelName string) ([]api.Tensor, error) {
-	manifest, err := imagegen.LoadManifest(modelName)
+func GetSafetensorsTensorInfo(name model.Name) ([]api.Tensor, error) {
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return nil, fmt.Errorf("failed to load manifest: %w", err)
 	}
 
-	return getTensorInfoFromManifest(manifest)
+	return getTensorInfoFromManifest(mf)
 }
 
 // getTensorInfoFromManifest extracts tensor info from a manifest.
 // This is separated for testability.
-func getTensorInfoFromManifest(manifest *imagegen.ModelManifest) ([]api.Tensor, error) {
+func getTensorInfoFromManifest(mf *manifest.Manifest) ([]api.Tensor, error) {
 	var tensors []api.Tensor
 
-	for _, layer := range manifest.Manifest.Layers {
-		if layer.MediaType != "application/vnd.ollama.image.tensor" {
+	for _, layer := range mf.Layers {
+		if layer.MediaType != manifest.MediaTypeImageTensor {
 			continue
 		}
 
 		// Read the safetensors header from the blob
-		blobPath := manifest.BlobPath(layer.Digest)
+		blobPath, err := manifest.BlobsPath(layer.Digest)
+		if err != nil {
+			continue
+		}
 		info, err := readSafetensorsHeader(blobPath)
 		if err != nil {
 			// Skip tensors we can't read
@@ -197,15 +201,15 @@ func getTensorInfoFromManifest(manifest *imagegen.ModelManifest) ([]api.Tensor,
 // GetSafetensorsDtype returns the quantization type for a safetensors model.
 // If the model is quantized (has _scale tensors), returns the quantization type (e.g., "FP8").
 // Otherwise returns the torch_dtype from config.json.
-func GetSafetensorsDtype(modelName string) (string, error) {
-	manifest, err := imagegen.LoadManifest(modelName)
+func GetSafetensorsDtype(name model.Name) (string, error) {
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return "", fmt.Errorf("failed to load manifest: %w", err)
 	}
 
 	// Check if model is quantized by looking for _scale tensors
-	for _, layer := range manifest.Manifest.Layers {
-		if layer.MediaType == "application/vnd.ollama.image.tensor" {
+	for _, layer := range mf.Layers {
+		if layer.MediaType == manifest.MediaTypeImageTensor {
 			if strings.HasSuffix(layer.Name, "_scale") {
 				// Model is quantized - return FP8 (affine quantization)
 				return "FP8", nil
@@ -217,7 +221,7 @@ func GetSafetensorsDtype(modelName string) (string, error) {
 	var cfg struct {
 		TorchDtype string `json:"torch_dtype"`
 	}
-	if err := manifest.ReadConfigJSON("config.json", &cfg); err != nil {
+	if err := mf.ReadConfigJSON("config.json", &cfg); err != nil {
 		return "", fmt.Errorf("failed to read config.json: %w", err)
 	}
 

x/server/show_test.go

@@ -8,7 +8,7 @@ import (
 	"path/filepath"
 	"testing"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/manifest"
 )
 
 func TestBuildModelInfo(t *testing.T) {
@@ -451,8 +451,14 @@ func TestParseSafetensorsHeader_Errors(t *testing.T) {
 }
 
 func TestGetTensorInfoFromManifest(t *testing.T) {
-	// Create a temp directory for blobs
+	// Create a temp directory for blobs and set OLLAMA_MODELS
 	tempDir := t.TempDir()
+	t.Setenv("OLLAMA_MODELS", tempDir)
+
+	blobDir := filepath.Join(tempDir, "blobs")
+	if err := os.MkdirAll(blobDir, 0o755); err != nil {
+		t.Fatalf("failed to create blobs dir: %v", err)
+	}
 
 	// Create test tensor blobs
 	tensors := []struct {
@@ -463,26 +469,26 @@ func TestGetTensorInfoFromManifest(t *testing.T) {
 	}{
 		{
 			name:   "model.embed_tokens.weight",
-			digest: "sha256:abc123",
+			digest: "sha256:abc123abc123abc123abc123abc123abc123abc123abc123abc123abc123abc0",
 			dtype:  "BF16",
 			shape:  []int64{262144, 2560},
 		},
 		{
 			name:   "model.layers.0.self_attn.q_proj.weight",
-			digest: "sha256:def456",
+			digest: "sha256:def456def456def456def456def456def456def456def456def456def456def0",
 			dtype:  "BF16",
 			shape:  []int64{2560, 2560},
 		},
 		{
 			name:   "model.norm.weight",
-			digest: "sha256:ghi789",
+			digest: "sha256:789789789789789789789789789789789789789789789789789789789789abc0",
 			dtype:  "F32",
 			shape:  []int64{2560},
 		},
 	}
 
 	// Create blob files
-	var layers []imagegen.ManifestLayer
+	var layers []manifest.Layer
 	for _, tensor := range tensors {
 		// Create safetensors blob
 		header := map[string]any{
@@ -498,15 +504,17 @@ func TestGetTensorInfoFromManifest(t *testing.T) {
 		binary.Write(&buf, binary.LittleEndian, uint64(len(headerJSON)))
 		buf.Write(headerJSON)
 
-		// Write blob file
-		blobName := "sha256-" + tensor.digest[7:]
-		blobPath := filepath.Join(tempDir, blobName)
+		// Write blob file using the digest format expected by GetBlobsPath
+		blobPath, err := manifest.BlobsPath(tensor.digest)
+		if err != nil {
+			t.Fatalf("failed to get blob path: %v", err)
+		}
 		if err := os.WriteFile(blobPath, buf.Bytes(), 0o644); err != nil {
 			t.Fatalf("failed to write blob: %v", err)
 		}
 
-		layers = append(layers, imagegen.ManifestLayer{
-			MediaType: "application/vnd.ollama.image.tensor",
+		layers = append(layers, manifest.Layer{
+			MediaType: manifest.MediaTypeImageTensor,
 			Digest:    tensor.digest,
 			Size:      int64(buf.Len() + 1000), // header + fake data
 			Name:      tensor.name,
@@ -514,21 +522,20 @@ func TestGetTensorInfoFromManifest(t *testing.T) {
 	}
 
 	// Add a non-tensor layer (should be skipped)
-	layers = append(layers, imagegen.ManifestLayer{
+	layers = append(layers, manifest.Layer{
 		MediaType: "application/vnd.ollama.image.json",
-		Digest:    "sha256:config",
+		Digest:    "sha256:0000000000000000000000000000000000000000000000000000000000000000",
 		Size:      100,
 		Name:      "config.json",
 	})
 
-	manifest := &imagegen.ModelManifest{
-		Manifest: &imagegen.Manifest{
-			Layers: layers,
-		},
-		BlobDir: tempDir,
+	mf := &manifest.Manifest{
+		SchemaVersion: 2,
+		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
+		Layers:        layers,
 	}
 
-	result, err := getTensorInfoFromManifest(manifest)
+	result, err := getTensorInfoFromManifest(mf)
 	if err != nil {
 		t.Fatalf("getTensorInfoFromManifest() error = %v", err)
 	}