diff --git a/backend/go/vllm-cpp/backend.go b/backend/go/vllm-cpp/backend.go
index 071b8f584..a53bf951d 100644
--- a/backend/go/vllm-cpp/backend.go
+++ b/backend/go/vllm-cpp/backend.go
@@ -109,6 +109,16 @@ func (v *VllmCpp) Load(opts *pb.ModelOptions) error {
 
 	v.opts = parseOptions(opts)
 
+	// A DFlash draft is a second checkpoint the engine opens by path, and the
+	// engine never downloads one. Resolve it against LocalAI's models directory
+	// now so a repo-id spelling works, and so a missing draft fails here with an
+	// actionable message rather than as an HF-cache miss inside the load.
+	resolvedSpec, err := resolveDraftModelPath(v.opts.speculativeConfig, opts.ModelPath)
+	if err != nil {
+		return err
+	}
+	v.opts.speculativeConfig = resolvedSpec
+
 	mp := defaultModelParams()
 	if v.opts.blockSize > 0 {
 		mp.BlockSize = v.opts.blockSize
diff --git a/backend/go/vllm-cpp/options.go b/backend/go/vllm-cpp/options.go
index b1fb2a4dd..b2d9a378e 100644
--- a/backend/go/vllm-cpp/options.go
+++ b/backend/go/vllm-cpp/options.go
@@ -18,6 +18,10 @@ package main
 
 import (
 	"encoding/json"
+	"fmt"
+	"os"
+	"path"
+	"path/filepath"
 	"strconv"
 	"strings"
 
@@ -206,6 +210,76 @@ func jsonInt32(v any, fallback int32) int32 {
 	}
 }
 
+// resolveDraftModelPath rewrites a DFlash draft reference into an absolute path
+// the engine can actually open.
+//
+// The engine resolves `speculative_config.model` against a directory containing
+// config.json, or against ~/.cache/huggingface/hub/models--<org>--<repo>/
+// snapshots/* - and it NEVER downloads. LocalAI keeps models in its own
+// directory, so a bare HF repo id (the spelling the vLLM docs teach) misses the
+// HF cache and dies deep in the load with "draft checkpoint not found", which
+// reads like a broken checkpoint rather than a missing download.
+//
+// So: try the reference as given, then the last path segment under the models
+// dir (`z-lab/Qwen3.6-27B-DFlash` -> `<models>/Qwen3.6-27B-DFlash`, which is
+// what LocalAI's own downloader produces), then the whole reference under the
+// models dir. If none exist, fail HERE with a message naming both what was
+// asked for and where we looked.
+//
+// mtp and ngram carry no separate draft checkpoint, so they pass through. A
+// document that does not parse also passes through: the engine owns config
+// validation and produces the better error.
+func resolveDraftModelPath(speculativeConfig, modelsDir string) (string, error) {
+	if strings.TrimSpace(speculativeConfig) == "" {
+		return speculativeConfig, nil
+	}
+	var spec map[string]any
+	if err := json.Unmarshal([]byte(speculativeConfig), &spec); err != nil {
+		return speculativeConfig, nil
+	}
+	if method, _ := spec["method"].(string); !strings.EqualFold(method, "dflash") {
+		return speculativeConfig, nil
+	}
+
+	ref, _ := spec["model"].(string)
+	ref = strings.TrimSpace(ref)
+	if ref == "" {
+		return "", fmt.Errorf(
+			"vllm-cpp: speculative_config method %q requires a \"model\" key naming the draft checkpoint", "dflash")
+	}
+
+	candidates := []string{ref}
+	if modelsDir != "" {
+		if base := path.Base(filepath.ToSlash(ref)); base != "" && base != "." && base != "/" {
+			candidates = append(candidates, filepath.Join(modelsDir, base))
+		}
+		candidates = append(candidates, filepath.Join(modelsDir, filepath.FromSlash(ref)))
+	}
+
+	for _, c := range candidates {
+		if _, err := os.Stat(filepath.Join(c, "config.json")); err != nil {
+			continue
+		}
+		abs, err := filepath.Abs(c)
+		if err != nil {
+			abs = c
+		}
+		spec["model"] = abs
+		out, err := json.Marshal(spec)
+		if err != nil {
+			return "", fmt.Errorf("vllm-cpp: re-encoding speculative_config: %w", err)
+		}
+		xlog.Info("[vllm-cpp] resolved DFlash draft checkpoint", "reference", ref, "path", abs)
+		return string(out), nil
+	}
+
+	return "", fmt.Errorf(
+		"vllm-cpp: DFlash draft checkpoint %q not found (looked in: %s). "+
+			"The engine does not download drafts - install the draft model into LocalAI first, "+
+			"or set speculative_config.model to an absolute path to a directory containing config.json",
+		ref, strings.Join(candidates, ", "))
+}
+
 func parseInt32(s string, fallback int32) int32 {
 	n, err := strconv.ParseInt(strings.TrimSpace(s), 10, 32)
 	if err != nil || n <= 0 {
diff --git a/backend/go/vllm-cpp/vllmcpp_test.go b/backend/go/vllm-cpp/vllmcpp_test.go
index 96112345d..126cff2c3 100644
--- a/backend/go/vllm-cpp/vllmcpp_test.go
+++ b/backend/go/vllm-cpp/vllmcpp_test.go
@@ -265,6 +265,91 @@ var _ = Describe("samplingFromPredict", func() {
 	})
 })
 
+// The engine resolves speculative_config.model against a local directory or
+// ~/.cache/huggingface/hub ONLY - it never downloads. LocalAI keeps models in
+// its own directory, so a bare repo id would miss the HF cache and fail deep in
+// the load with a confusing "draft checkpoint not found". Resolve it here.
+var _ = Describe("resolveDraftModelPath", func() {
+	var modelsDir string
+
+	BeforeEach(func() {
+		modelsDir = GinkgoT().TempDir()
+	})
+
+	// draftDir creates a plausible draft checkpoint under models/.
+	draftDir := func(name string) string {
+		d := filepath.Join(modelsDir, name)
+		Expect(os.MkdirAll(d, 0o750)).To(Succeed())
+		Expect(os.WriteFile(filepath.Join(d, "config.json"), []byte("{}"), 0o600)).To(Succeed())
+		return d
+	}
+
+	It("rewrites a repo id to the matching directory in the models dir", func() {
+		want := draftDir("Qwen3.6-27B-DFlash")
+		spec := `{"method":"dflash","model":"z-lab/Qwen3.6-27B-DFlash"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(MatchJSON(`{"method":"dflash","model":"` + want + `"}`))
+	})
+
+	It("rewrites a models-dir-relative path", func() {
+		want := draftDir("drafts__dflash")
+		spec := `{"method":"dflash","model":"drafts__dflash"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(ContainSubstring(want))
+	})
+
+	It("leaves an absolute path that already resolves alone", func() {
+		abs := draftDir("elsewhere")
+		spec := `{"method":"dflash","model":"` + abs + `"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(MatchJSON(spec))
+	})
+
+	It("fails with an actionable error when the draft is nowhere on disk", func() {
+		// Silently passing the repo id through would surface as an HF-cache
+		// miss inside the engine, which reads as "your model is broken".
+		spec := `{"method":"dflash","model":"z-lab/Not-Downloaded"}`
+		_, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("z-lab/Not-Downloaded"))
+		Expect(err.Error()).To(ContainSubstring(modelsDir))
+	})
+
+	It("requires a model key for dflash", func() {
+		_, err := resolveDraftModelPath(`{"method":"dflash"}`, modelsDir)
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("model"))
+	})
+
+	It("leaves mtp and ngram configs untouched", func() {
+		// Neither has a separate draft checkpoint to resolve.
+		for _, spec := range []string{
+			`{"method":"mtp"}`,
+			`{"method":"ngram","num_speculative_tokens":4}`,
+		} {
+			out, err := resolveDraftModelPath(spec, modelsDir)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(out).To(MatchJSON(spec))
+		}
+	})
+
+	It("passes a malformed document through for the engine to reject", func() {
+		// The engine owns config validation and produces the better message.
+		out, err := resolveDraftModelPath(`{not json`, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(Equal(`{not json`))
+	})
+
+	It("is a no-op on an empty config", func() {
+		out, err := resolveDraftModelPath("", modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(BeEmpty())
+	})
+})
+
 var _ = Describe("validModelPath", func() {
 	It("accepts a .gguf file", func() {
 		dir := GinkgoT().TempDir()
diff --git a/docs/content/features/text-generation.md b/docs/content/features/text-generation.md
index 16060420e..8c44a1e4c 100644
--- a/docs/content/features/text-generation.md
+++ b/docs/content/features/text-generation.md
@@ -962,7 +962,22 @@ prefill from blowing up the per-step activation on the hybrid architectures.
 #### Speculative decoding
 
 `speculative_config:` takes the same JSON object as vLLM's
-`--speculative-config`. Three methods are supported:
+`--speculative-config`. Three methods are supported.
+
+> **Architecture limit.** At the current engine pin, `mtp` and `dflash` are
+> **Qwen3.5 / Qwen3.6 only**. The engine builds a widened speculative KV cache
+> directly for those families rather than through the model registry, so a
+> speculative config on any other architecture (Llama, GLM, Gemma, Mistral, ...)
+> will not work regardless of checkpoint format. `ngram` needs no draft weights
+> and is not subject to this limit.
+
+> **Format limit.** `mtp` and `dflash` require a **safetensors** target and are
+> rejected at load on a `.gguf` target, with:
+> `speculative decoding requires a safetensors target checkpoint`.
+> This is a current gap in the engine's GGUF loader, not a property of the GGUF
+> format - GGUF can carry MTP weights (llama.cpp reads them as `nextn.*`
+> tensors plus a `<arch>.nextn_predict_layers` key), but vllm.cpp's GGUF path
+> does not map them yet. `ngram` works fine on GGUF.
 
 **MTP** (Multi-Token Prediction) uses a draft head shipped inside the target
 checkpoint's own `mtp.*` tensors, so there is no second model to download. It
@@ -990,6 +1005,18 @@ engine_args:
     num_speculative_tokens: 4
 ```
 
+The draft shares the *target's* `embed_tokens` and `lm_head`, so both must come
+from the same model family and the target must be safetensors.
+
+**The engine does not download the draft.** `model:` is resolved, in order,
+as a path as given, then as the last path segment under LocalAI's models
+directory (`z-lab/Qwen3.6-27B-DFlash` → `<models>/Qwen3.6-27B-DFlash`, which is
+what LocalAI's own downloader produces), then as the whole reference under the
+models directory. Install the draft into LocalAI first, or give an absolute path
+to a directory containing `config.json`. If none of those resolve, the load
+fails immediately naming every location that was tried, rather than reporting a
+missing checkpoint from inside the engine.
+
 **N-gram** needs no draft model at all - it proposes from the prompt's own
 suffix history. `num_speculative_tokens` is required: