feat: generic chat_template_kwargs (model config + per-request metadata) (#10359)

* feat(config): add chat_template_kwargs model field + resolver

Adds the ChatTemplateKwargs model-config map and RequestMetadata carrier,
plus ResolveChatTemplateKwargs which layers the config map under coerced
request metadata. Foundation for generic jinja chat-template kwargs (issue #10329).

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(backend): forward resolved chat_template_kwargs blob to backends

gRPCPredictOpts now merges per-request client metadata over the server-derived
enable_thinking/reasoning_effort (reaching all backends via the standalone keys)
and serialises the resolved chat_template_kwargs map into a JSON blob for
llama.cpp, written last so a client cannot clobber it. Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(http): wire request metadata to config.RequestMetadata

The OpenAI request metadata field was parsed but unused; stamp it onto the
per-request ModelConfig so gRPCPredictOpts forwards it as chat_template_kwargs
overrides. Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(llama-cpp): generic chat_template_kwargs merge (drop per-key blocks)

Replace the per-key enable_thinking/reasoning_effort handling in both the
streaming and non-streaming chat paths with a single block that parses the
chat_template_kwargs JSON blob resolved by the Go layer and merges every key
into body_json. New jinja template levers (e.g. preserve_thinking) now need
no C++ change. Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* docs: document custom chat_template_kwargs (model + per-request)

Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* test(backend): pin reasoning_effort as a string in the chat_template_kwargs blob

Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* test(http): e2e guard pinning chat_template_kwargs forwarded to gRPC

Adds an ECHO_PREDICT_METADATA marker to the mock-backend that echoes the
received PredictOptions.Metadata, and an app_test.go spec that drives a real
/v1/chat/completions request (model chat_template_kwargs + per-request metadata
override) and asserts the exact metadata + chat_template_kwargs blob the REST
layer forwards to gRPC. Locks the REST->gRPC contract against regressions. Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* test(config): grandfather chat_template_kwargs in registry coverage

chat_template_kwargs is a free-form map[string]any (like engine_args, already
on the list), not a scalar the config UI registry can surface, so it is exempt
from the registry-entry requirement. Fixes the TestAllFieldsHaveRegistryEntries
failure introduced by the new field. Issue #10329.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>

backend/cpp/llama-cpp/grpc-server.cpp

@@ -1922,25 +1922,27 @@ public:
                     body_json["min_p"] = data["min_p"];
                 }
 
-                // Pass enable_thinking via chat_template_kwargs (where oaicompat_chat_params_parse reads it)
+                // Forward the chat_template_kwargs the Go layer resolved (model config
+                // chat_template_kwargs + per-request metadata: enable_thinking,
+                // reasoning_effort, preserve_thinking, ...). One generic merge replaces
+                // the previous per-key handling - new template levers need no C++ change.
+                // oaicompat_chat_params_parse reads these from body_json.
                 const auto& metadata = request->metadata();
-                auto et_it = metadata.find("enable_thinking");
-                if (et_it != metadata.end()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
+                auto ctk_it = metadata.find("chat_template_kwargs");
+                if (ctk_it != metadata.end() && !ctk_it->second.empty()) {
+                    try {
+                        json ctk = json::parse(ctk_it->second);
+                        if (ctk.is_object()) {
+                            if (!body_json.contains("chat_template_kwargs")) {
+                                body_json["chat_template_kwargs"] = json::object();
+                            }
+                            for (auto& el : ctk.items()) {
+                                body_json["chat_template_kwargs"][el.key()] = el.value();
+                            }
+                        }
+                    } catch (const std::exception & e) {
+                        SRV_WRN("failed to parse chat_template_kwargs metadata: %s\n", e.what());
                     }
-                    body_json["chat_template_kwargs"]["enable_thinking"] = (et_it->second == "true");
-                }
-
-                // Pass reasoning_effort via chat_template_kwargs too: the lever
-                // jinja templates like gpt-oss (Harmony) / LFM2.5 read, distinct
-                // from enable_thinking which those templates ignore.
-                auto re_it = metadata.find("reasoning_effort");
-                if (re_it != metadata.end() && !re_it->second.empty()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
-                    }
-                    body_json["chat_template_kwargs"]["reasoning_effort"] = re_it->second;
                 }
 
                 // Debug: Print full body_json before template processing (includes messages, tools, tool_choice, etc.)
@@ -2756,25 +2758,26 @@ public:
                     body_json["min_p"] = data["min_p"];
                 }
 
-                // Pass enable_thinking via chat_template_kwargs (where oaicompat_chat_params_parse reads it)
+                // Forward the chat_template_kwargs the Go layer resolved (model config
+                // chat_template_kwargs + per-request metadata: enable_thinking,
+                // reasoning_effort, preserve_thinking, ...). One generic merge replaces
+                // the previous per-key handling - new template levers need no C++ change.
                 const auto& predict_metadata = request->metadata();
-                auto predict_et_it = predict_metadata.find("enable_thinking");
-                if (predict_et_it != predict_metadata.end()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
+                auto predict_ctk_it = predict_metadata.find("chat_template_kwargs");
+                if (predict_ctk_it != predict_metadata.end() && !predict_ctk_it->second.empty()) {
+                    try {
+                        json ctk = json::parse(predict_ctk_it->second);
+                        if (ctk.is_object()) {
+                            if (!body_json.contains("chat_template_kwargs")) {
+                                body_json["chat_template_kwargs"] = json::object();
+                            }
+                            for (auto& el : ctk.items()) {
+                                body_json["chat_template_kwargs"][el.key()] = el.value();
+                            }
+                        }
+                    } catch (const std::exception & e) {
+                        SRV_WRN("failed to parse chat_template_kwargs metadata: %s\n", e.what());
                     }
-                    body_json["chat_template_kwargs"]["enable_thinking"] = (predict_et_it->second == "true");
-                }
-
-                // Pass reasoning_effort via chat_template_kwargs too: the lever
-                // jinja templates like gpt-oss (Harmony) / LFM2.5 read, distinct
-                // from enable_thinking which those templates ignore.
-                auto predict_re_it = predict_metadata.find("reasoning_effort");
-                if (predict_re_it != predict_metadata.end() && !predict_re_it->second.empty()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
-                    }
-                    body_json["chat_template_kwargs"]["reasoning_effort"] = predict_re_it->second;
                 }
 
                 // Debug: Print full body_json before template processing (includes messages, tools, tool_choice, etc.)

core/backend/options.go

@@ -368,6 +368,25 @@ func gRPCPredictOpts(c config.ModelConfig, modelPath string) *pb.PredictOptions
 	if c.ReasoningEffort != "" {
 		metadata["reasoning_effort"] = c.ReasoningEffort
 	}
+	// Client request metadata overrides the server-derived reasoning levers and
+	// reaches every backend through these standalone string keys (Python backends
+	// read them directly). The reserved blob key is server-owned and skipped.
+	for k, v := range c.RequestMetadata {
+		if k == "chat_template_kwargs" {
+			continue
+		}
+		metadata[k] = v
+	}
+	// Build the generic chat_template_kwargs blob (model config map + coerced
+	// metadata) for llama.cpp and write it LAST so a client cannot clobber it.
+	if blob := c.ResolveChatTemplateKwargs(metadata); len(blob) > 0 {
+		b, err := json.Marshal(blob)
+		if err != nil {
+			xlog.Warn("failed to marshal chat_template_kwargs", "error", err)
+		} else {
+			metadata["chat_template_kwargs"] = string(b)
+		}
+	}
 	pbOpts.Metadata = metadata
 
 	// Logprobs and TopLogprobs are set by the caller if provided

core/backend/options_internal_test.go

@@ -161,3 +161,67 @@ var _ = Describe("grpcModelOpts NBatch", func() {
 		Expect(opts.ContextSize).To(BeEquivalentTo(4096), "n_batch must match the effective n_ctx the backend receives")
 	})
 })
+
+// Guards the generic chat_template_kwargs forwarding: the model config map plus any
+// per-request metadata overrides are merged, coerced, and serialised into the
+// backend metadata blob that llama.cpp reads. Client metadata also overrides the
+// server-derived standalone enable_thinking key (cross-backend consistency).
+var _ = Describe("gRPCPredictOpts chat_template_kwargs metadata", func() {
+	baseCfg := func() config.ModelConfig {
+		cfg := config.ModelConfig{}
+		cfg.SetDefaults()
+		return cfg
+	}
+
+	It("serialises the config map into the chat_template_kwargs blob", func() {
+		cfg := baseCfg()
+		cfg.ChatTemplateKwargs = map[string]any{"preserve_thinking": true}
+		opts := gRPCPredictOpts(cfg, "/tmp/models")
+		Expect(opts.Metadata).To(HaveKey("chat_template_kwargs"))
+		var blob map[string]any
+		Expect(json.Unmarshal([]byte(opts.Metadata["chat_template_kwargs"]), &blob)).To(Succeed())
+		Expect(blob).To(HaveKeyWithValue("preserve_thinking", true))
+	})
+
+	It("serialises reasoning_effort into the blob as a JSON string", func() {
+		cfg := baseCfg()
+		cfg.ReasoningEffort = "high"
+		opts := gRPCPredictOpts(cfg, "/tmp/models")
+		Expect(opts.Metadata).To(HaveKey("chat_template_kwargs"))
+		var blob map[string]any
+		Expect(json.Unmarshal([]byte(opts.Metadata["chat_template_kwargs"]), &blob)).To(Succeed())
+		// reasoning_effort must remain a string in the blob (jinja templates that
+		// key on the level read a string), unlike enable_thinking which is a bool.
+		Expect(blob["reasoning_effort"]).To(BeAssignableToTypeOf(""))
+		Expect(blob).To(HaveKeyWithValue("reasoning_effort", "high"))
+	})
+
+	It("lets client request metadata override the server-derived enable_thinking key", func() {
+		cfg := baseCfg()
+		disable := true
+		cfg.ReasoningConfig = reasoning.Config{DisableReasoning: &disable} // server: enable_thinking=false
+		cfg.RequestMetadata = map[string]string{"enable_thinking": "true"} // client overrides
+		opts := gRPCPredictOpts(cfg, "/tmp/models")
+		// standalone key (Python backends) reflects the client override
+		Expect(opts.Metadata).To(HaveKeyWithValue("enable_thinking", "true"))
+		// blob (llama.cpp) reflects it too, as a real bool
+		var blob map[string]any
+		Expect(json.Unmarshal([]byte(opts.Metadata["chat_template_kwargs"]), &blob)).To(Succeed())
+		Expect(blob).To(HaveKeyWithValue("enable_thinking", true))
+	})
+
+	It("does not let a client clobber the blob via a chat_template_kwargs metadata key", func() {
+		cfg := baseCfg()
+		cfg.ChatTemplateKwargs = map[string]any{"preserve_thinking": true}
+		cfg.RequestMetadata = map[string]string{"chat_template_kwargs": "{\"preserve_thinking\": false}"}
+		opts := gRPCPredictOpts(cfg, "/tmp/models")
+		var blob map[string]any
+		Expect(json.Unmarshal([]byte(opts.Metadata["chat_template_kwargs"]), &blob)).To(Succeed())
+		Expect(blob).To(HaveKeyWithValue("preserve_thinking", true))
+	})
+
+	It("omits the blob when there is nothing to forward", func() {
+		opts := gRPCPredictOpts(baseCfg(), "/tmp/models")
+		Expect(opts.Metadata).ToNot(HaveKey("chat_template_kwargs"))
+	})
+})

core/config/chat_template_kwargs_test.go

@@ -0,0 +1,48 @@
+package config_test
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+)
+
+// ResolveChatTemplateKwargs layers the model config map (base) under the coerced
+// backend metadata (server reasoning levers + client request overrides).
+var _ = Describe("ModelConfig.ResolveChatTemplateKwargs", func() {
+	It("returns nil when nothing is set", func() {
+		c := &config.ModelConfig{}
+		Expect(c.ResolveChatTemplateKwargs(nil)).To(BeNil())
+	})
+
+	It("returns the config map when no metadata is present", func() {
+		c := &config.ModelConfig{ChatTemplateKwargs: map[string]any{"preserve_thinking": true}}
+		Expect(c.ResolveChatTemplateKwargs(nil)).To(HaveKeyWithValue("preserve_thinking", true))
+	})
+
+	It("lets metadata override the config map", func() {
+		c := &config.ModelConfig{ChatTemplateKwargs: map[string]any{"enable_thinking": true}}
+		got := c.ResolveChatTemplateKwargs(map[string]string{"enable_thinking": "false"})
+		Expect(got).To(HaveKeyWithValue("enable_thinking", false))
+	})
+
+	It("coerces true/false to bool and leaves other strings as-is", func() {
+		c := &config.ModelConfig{}
+		got := c.ResolveChatTemplateKwargs(map[string]string{
+			"enable_thinking":  "true",
+			"reasoning_effort": "high",
+		})
+		Expect(got).To(HaveKeyWithValue("enable_thinking", true))
+		Expect(got).To(HaveKeyWithValue("reasoning_effort", "high"))
+	})
+
+	It("skips the reserved chat_template_kwargs metadata key but keeps siblings", func() {
+		c := &config.ModelConfig{}
+		got := c.ResolveChatTemplateKwargs(map[string]string{
+			"chat_template_kwargs": "{\"x\":1}",
+			"preserve_thinking":    "true",
+		})
+		Expect(got).ToNot(HaveKey("chat_template_kwargs"))
+		Expect(got).To(HaveKeyWithValue("preserve_thinking", true))
+	})
+})

core/config/meta/registry_coverage_test.go

@@ -112,6 +112,7 @@ var grandfatheredUnregistered = []string{
 	"agent.max_attempts",
 	"agent.max_iterations",
 	"cfg_scale",
+	"chat_template_kwargs",
 	"concurrency_groups",
 	"cutstrings",
 	"debug",

core/config/model_config.go

@@ -70,6 +70,19 @@ type ModelConfig struct {
 	// (Harmony) or LFM2.5 — honor it; "none" also toggles enable_thinking off.
 	ReasoningEffort string `yaml:"reasoning_effort,omitempty" json:"reasoning_effort,omitempty"`
 
+	// ChatTemplateKwargs are arbitrary key/values forwarded to the backend's jinja
+	// chat template via chat_template_kwargs (e.g. preserve_thinking: true). The
+	// server-derived reasoning levers (enable_thinking / reasoning_effort) and any
+	// per-request metadata overrides layer on top. See gRPCPredictOpts.
+	ChatTemplateKwargs map[string]any `yaml:"chat_template_kwargs,omitempty" json:"chat_template_kwargs,omitempty"`
+
+	// RequestMetadata holds the raw client request `metadata` map for the current
+	// request. The request middleware stamps it; gRPCPredictOpts merges it into the
+	// backend gRPC metadata (overriding the server-derived enable_thinking /
+	// reasoning_effort) and folds it, coerced, into the chat_template_kwargs blob.
+	// Never persisted to YAML.
+	RequestMetadata map[string]string `yaml:"-" json:"-"`
+
 	FeatureFlag FeatureFlag `yaml:"feature_flags,omitempty" json:"feature_flags,omitempty"` // Feature Flag registry. We move fast, and features may break on a per model/backend basis. Registry for (usually temporary) flags that indicate aborting something early.
 	// LLM configs (GPT4ALL, Llama.cpp, ...)
 	LLMConfig `yaml:",inline" json:",inline"`
@@ -551,6 +564,44 @@ func (c *ModelConfig) ApplyReasoningEffort(requestEffort string) {
 	}
 }
 
+// coerceChatTemplateKwarg coerces a request-metadata string value for use as a
+// jinja chat_template_kwarg. "true"/"false" become real booleans (so a jinja
+// `{% if preserve_thinking %}` reads false correctly, since any non-empty string
+// is truthy); everything else stays a string. Numeric/typed per-request values are
+// out of scope - set those in the model YAML chat_template_kwargs (YAML keeps the type).
+func coerceChatTemplateKwarg(v string) any {
+	switch v {
+	case "true":
+		return true
+	case "false":
+		return false
+	default:
+		return v
+	}
+}
+
+// ResolveChatTemplateKwargs builds the final chat_template_kwargs map forwarded to
+// the backend, layered: the model config map (base) < the coerced backend metadata
+// (server reasoning levers + client request overrides). `meta` is the already-merged
+// backend metadata string map. The reserved "chat_template_kwargs" key is skipped so
+// a client cannot smuggle a nested blob. Returns nil when there is nothing to forward.
+func (c *ModelConfig) ResolveChatTemplateKwargs(meta map[string]string) map[string]any {
+	out := map[string]any{}
+	for k, v := range c.ChatTemplateKwargs {
+		out[k] = v
+	}
+	for k, v := range meta {
+		if k == "chat_template_kwargs" {
+			continue
+		}
+		out[k] = coerceChatTemplateKwarg(v)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
 // @Description PipelineStreaming toggles incremental delivery per realtime stage.
 type PipelineStreaming struct {
 	LLM           *bool `yaml:"llm,omitempty" json:"llm,omitempty"`

core/http/app_test.go

@@ -735,6 +735,18 @@ parameters:
 `
 			Expect(os.WriteFile(filepath.Join(modelDir, "mock-model.yaml"), []byte(mockModelYAML), 0644)).To(Succeed())
 
+			// A second model carrying chat_template_kwargs so the REST->gRPC
+			// metadata-forwarding spec below can assert the model-YAML kwarg is
+			// merged with the per-request override.
+			mockCTKModelYAML := `name: mock-ctk-model
+backend: mock-backend
+parameters:
+  model: mock-model.bin
+chat_template_kwargs:
+  preserve_thinking: true
+`
+			Expect(os.WriteFile(filepath.Join(modelDir, "mock-ctk-model.yaml"), []byte(mockCTKModelYAML), 0644)).To(Succeed())
+
 			systemState, err := system.GetSystemState(
 				system.WithBackendPath(backendDir),
 				system.WithModelPath(modelDir),
@@ -809,6 +821,59 @@ parameters:
 			Expect(string(dat)).To(ContainSubstring("mock-backend"))
 		})
 
+		It("forwards chat_template_kwargs and reasoning levers to gRPC PredictOptions.Metadata", func() {
+			// True HTTP->gRPC contract guard: drive a real /v1/chat/completions
+			// request and assert the exact metadata the REST layer forwarded to
+			// the backend. The mock-backend echoes PredictOptions.Metadata as JSON
+			// when it sees the ECHO_PREDICT_METADATA marker in the prompt, so this
+			// pins the request->gRPC mapping (model-YAML chat_template_kwargs +
+			// per-request metadata override + type coercion + standalone keys)
+			// without adding a new RPC. The marker rides in the user content and
+			// must survive into the backend prompt; if a future default chat
+			// template drops raw user content, move the marker to /v1/completions.
+			reqBody := map[string]any{
+				"model": "mock-ctk-model",
+				"messages": []map[string]any{
+					{"role": "user", "content": "ECHO_PREDICT_METADATA"},
+				},
+				// per-request override: overrides the standalone enable_thinking key
+				// and exercises coercion ("false" -> bool, "low" -> string) in the blob
+				"metadata": map[string]string{
+					"enable_thinking":  "false",
+					"reasoning_effort": "low",
+				},
+			}
+
+			var chatResp struct {
+				Choices []struct {
+					Message struct {
+						Content string `json:"content"`
+					} `json:"message"`
+				} `json:"choices"`
+			}
+			err := postRequestResponseJSON("http://127.0.0.1:9090/v1/chat/completions", &reqBody, &chatResp)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(chatResp.Choices).ToNot(BeEmpty())
+
+			// The assistant content is the JSON snapshot of PredictOptions.Metadata.
+			var meta map[string]string
+			Expect(json.Unmarshal([]byte(chatResp.Choices[0].Message.Content), &meta)).To(Succeed(), "echoed metadata: %s", chatResp.Choices[0].Message.Content)
+
+			// Standalone keys reflect the per-request override (consumed by Python
+			// backends; consistent across backends).
+			Expect(meta).To(HaveKeyWithValue("enable_thinking", "false"))
+			Expect(meta).To(HaveKeyWithValue("reasoning_effort", "low"))
+
+			// The chat_template_kwargs blob (consumed by llama.cpp) merges the
+			// model-YAML kwarg with the coerced request metadata override.
+			Expect(meta).To(HaveKey("chat_template_kwargs"))
+			var ctk map[string]any
+			Expect(json.Unmarshal([]byte(meta["chat_template_kwargs"]), &ctk)).To(Succeed(), "chat_template_kwargs blob: %s", meta["chat_template_kwargs"])
+			Expect(ctk).To(HaveKeyWithValue("preserve_thinking", true)) // bool from model YAML
+			Expect(ctk).To(HaveKeyWithValue("enable_thinking", false))  // coerced "false" -> bool
+			Expect(ctk).To(HaveKeyWithValue("reasoning_effort", "low")) // non-bool stays string
+		})
+
 		// Agent Jobs: HTTP API for task/job scheduling. The underlying AgentPool
 		// service is exercised in core/services/agentpool/agent_jobs_test.go;
 		// these specs cover the /api/agent/* HTTP plumbing on top.

core/http/middleware/request.go

@@ -318,6 +318,13 @@ func mergeOpenAIRequestAndModelConfig(config *config.ModelConfig, input *schema.
 	// (an operator's explicit disable wins over a request asking to think).
 	config.ApplyReasoningEffort(input.ReasoningEffort)
 
+	// Forward the client's request metadata so chat-template kwargs set per-request
+	// (enable_thinking, reasoning_effort, preserve_thinking, ...) reach the backend
+	// and override the model's reasoning-config defaults. See gRPCPredictOpts.
+	if len(input.Metadata) > 0 {
+		config.RequestMetadata = input.Metadata
+	}
+
 	// Collapse the modern max_completion_tokens alias into the
 	// legacy Maxtokens field so downstream code reads exactly one.
 	// MaxCompletionTokens wins on conflict — it's the canonical

core/http/middleware/request_test.go

@@ -731,3 +731,60 @@ var _ = Describe("SetModelAndConfig reasoning_effort parsing (chat completions)"
 		Expect(*(*captured2).ReasoningConfig.DisableReasoning).To(BeFalse())
 	})
 })
+
+var _ = Describe("SetModelAndConfig metadata passthrough (chat completions)", func() {
+	var modelDir string
+
+	BeforeEach(func() {
+		var err error
+		modelDir, err = os.MkdirTemp("", "localai-test-models-*")
+		Expect(err).ToNot(HaveOccurred())
+	})
+	AfterEach(func() { _ = os.RemoveAll(modelDir) })
+
+	buildApp := func() (*echo.Echo, **config.ModelConfig) {
+		Expect(os.WriteFile(filepath.Join(modelDir, "test-model.yaml"),
+			[]byte("name: test-model\nbackend: llama\n"), 0644)).To(Succeed())
+		ss := &system.SystemState{Model: system.Model{ModelsPath: modelDir}}
+		appConfig := config.NewApplicationConfig()
+		appConfig.SystemState = ss
+		mcl := config.NewModelConfigLoader(modelDir)
+		ml := model.NewModelLoader(ss)
+		re := NewRequestExtractor(mcl, ml, appConfig)
+
+		captured := new(*config.ModelConfig)
+		app := echo.New()
+		app.POST("/v1/chat/completions",
+			func(c echo.Context) error {
+				if cfg, ok := c.Get(CONTEXT_LOCALS_KEY_MODEL_CONFIG).(*config.ModelConfig); ok {
+					*captured = cfg
+				}
+				return c.String(http.StatusOK, "ok")
+			},
+			re.SetModelAndConfig(func() schema.LocalAIRequest { return new(schema.OpenAIRequest) }),
+			func(next echo.HandlerFunc) echo.HandlerFunc {
+				return func(c echo.Context) error {
+					if err := re.SetOpenAIRequest(c); err != nil {
+						return err
+					}
+					return next(c)
+				}
+			},
+		)
+		return app, captured
+	}
+
+	It("stamps request metadata onto the config", func() {
+		app, captured := buildApp()
+		body := `{"model":"test-model","messages":[{"role":"user","content":"hi"}],` +
+			`"metadata":{"preserve_thinking":"true"}}`
+		req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(body))
+		req.Header.Set("Content-Type", "application/json")
+		rec := httptest.NewRecorder()
+		app.ServeHTTP(rec, req)
+
+		Expect(rec.Code).To(Equal(http.StatusOK))
+		Expect(*captured).ToNot(BeNil())
+		Expect((*captured).RequestMetadata).To(HaveKeyWithValue("preserve_thinking", "true"))
+	})
+})

docs/content/advanced/model-configuration.md

@@ -438,6 +438,36 @@ pipeline:
   reasoning_effort: none   # overrides the LLM model's own reasoning_effort
 ```
 
+#### Custom `chat_template_kwargs`
+
+Some jinja chat templates expose extra variables beyond `enable_thinking` /
+`reasoning_effort` (for example Qwen3's `preserve_thinking`). Set arbitrary key/values in
+the model config and they are forwarded to the backend's `chat_template_kwargs` as-is, so
+you don't need a dedicated server option per template variable:
+
+```yaml
+name: qwen3
+chat_template_kwargs:
+  preserve_thinking: true
+```
+
+You can also override (or add) any of these per request through the OpenAI `metadata`
+field on `/v1/chat/completions`. Values are strings; `"true"` / `"false"` are coerced to
+booleans, anything else is passed through as a string:
+
+```json
+{
+  "model": "qwen3",
+  "messages": [{"role": "user", "content": "hi"}],
+  "metadata": { "preserve_thinking": "true", "enable_thinking": "false" }
+}
+```
+
+Per-request `metadata` overrides the model config defaults and the reasoning-config levers,
+and (for `enable_thinking` / `reasoning_effort`) takes effect across every backend that
+reads them, not just llama.cpp. Typed (non-boolean) values are only supported through the
+model YAML `chat_template_kwargs`, where YAML preserves the type.
+
 ### Multimodal Backend Options
 
 | Option | Type | Default | Description |

tests/e2e/mock-backend/main.go

@@ -109,6 +109,23 @@ func (m *MockBackend) Predict(ctx context.Context, in *pb.PredictOptions) (*pb.R
 		}, nil
 	}
 
+	// ECHO_PREDICT_METADATA lets tests assert exactly what the REST layer
+	// forwarded to the backend as gRPC PredictOptions.Metadata (e.g. the
+	// chat_template_kwargs blob and the standalone enable_thinking/reasoning_effort
+	// keys). The reply carries a JSON snapshot of in.Metadata so an HTTP-level
+	// test can pin the request -> gRPC mapping without a new RPC.
+	if strings.Contains(in.Prompt, "ECHO_PREDICT_METADATA") {
+		payload, err := json.Marshal(in.Metadata)
+		if err != nil {
+			return nil, fmt.Errorf("mock backend echo metadata error: %w", err)
+		}
+		return &pb.Reply{
+			Message:      payload,
+			Tokens:       int32(len(in.Metadata)),
+			PromptTokens: 1,
+		}, nil
+	}
+
 	// ECHO_SERVED_MODEL returns the loaded model file path so router e2e
 	// tests can verify which candidate actually served the request without
 	// adding a new RPC. The router fans out to a single backend process per