diff --git a/core/http/middleware/request.go b/core/http/middleware/request.go
index 8bddd002a5..6d4d851038 100644
--- a/core/http/middleware/request.go
+++ b/core/http/middleware/request.go
@@ -391,6 +391,40 @@ func extractToolChoiceFunctionName(m map[string]any) string {
 	return ""
 }
 
+// mergeChatTemplateKwargs folds a request's chat_template_kwargs (the
+// llama-server / vLLM / SGLang field) into its metadata map, so clients written
+// for those servers can set e.g. enable_thinking per request. Scalars become the
+// strings the metadata path expects (booleans as "true"/"false", which
+// ResolveChatTemplateKwargs turns back into booleans); nested objects and arrays
+// are skipped, as is the reserved "chat_template_kwargs" key. An explicit
+// metadata entry wins over the same key here. With no kwargs, metadata is
+// returned unchanged.
+func mergeChatTemplateKwargs(kwargs map[string]any, metadata map[string]string) map[string]string {
+	if len(kwargs) == 0 {
+		return metadata
+	}
+	out := make(map[string]string, len(kwargs)+len(metadata))
+	for k, v := range kwargs {
+		if k == "chat_template_kwargs" {
+			continue
+		}
+		switch t := v.(type) {
+		case bool:
+			out[k] = strconv.FormatBool(t)
+		case string:
+			out[k] = t
+		case float64:
+			out[k] = strconv.FormatFloat(t, 'f', -1, 64)
+		case json.Number:
+			out[k] = t.String()
+		}
+	}
+	for k, v := range metadata {
+		out[k] = v
+	}
+	return out
+}
+
 func mergeOpenAIRequestAndModelConfig(config *config.ModelConfig, input *schema.OpenAIRequest) error {
 	// Discard client tools before anything below derives functions or a forced
 	// tool_choice from them; the model's own MCP tools are injected later.
@@ -456,8 +490,9 @@ func mergeOpenAIRequestAndModelConfig(config *config.ModelConfig, input *schema.
 	// 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
+	// The llama-server style chat_template_kwargs field takes the same path.
+	if md := mergeChatTemplateKwargs(input.ChatTemplateKwargs, input.Metadata); len(md) > 0 {
+		config.RequestMetadata = md
 	}
 
 	// Collapse the modern max_completion_tokens alias into the
diff --git a/core/http/middleware/request_test.go b/core/http/middleware/request_test.go
index d2d1d539df..2d24a14e1a 100644
--- a/core/http/middleware/request_test.go
+++ b/core/http/middleware/request_test.go
@@ -917,6 +917,41 @@ var _ = Describe("SetModelAndConfig metadata passthrough (chat completions)", fu
 		Expect(*captured).ToNot(BeNil())
 		Expect((*captured).RequestMetadata).To(HaveKeyWithValue("preserve_thinking", "true"))
 	})
+
+	It("folds llama-server style chat_template_kwargs into the request metadata", func() {
+		app, captured := buildApp()
+		body := `{"model":"test-model","messages":[{"role":"user","content":"hi"}],` +
+			`"chat_template_kwargs":{"enable_thinking":false,"reasoning_effort":"low","budget":512,` +
+			`"nested":{"a":1},"list":[1],"chat_template_kwargs":"x"}}`
+		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(Equal(map[string]string{
+			"enable_thinking":  "false",
+			"reasoning_effort": "low",
+			"budget":           "512",
+		}))
+	})
+
+	It("lets an explicit metadata key win over chat_template_kwargs", func() {
+		app, captured := buildApp()
+		body := `{"model":"test-model","messages":[{"role":"user","content":"hi"}],` +
+			`"chat_template_kwargs":{"enable_thinking":false,"preserve_thinking":true},` +
+			`"metadata":{"enable_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("enable_thinking", "true"))
+		Expect((*captured).RequestMetadata).To(HaveKeyWithValue("preserve_thinking", "true"))
+	})
 })
 
 // A model config may default to decayed_mean pooling with a half-life
diff --git a/core/schema/openai.go b/core/schema/openai.go
index 62b2452621..2df9f9e8d9 100644
--- a/core/schema/openai.go
+++ b/core/schema/openai.go
@@ -263,6 +263,12 @@ type OpenAIRequest struct {
 	ReasoningEffort string `json:"reasoning_effort,omitempty" yaml:"reasoning_effort"`
 
 	Metadata map[string]string `json:"metadata,omitempty" yaml:"metadata"`
+
+	// ChatTemplateKwargs is the per-request chat_template_kwargs field that
+	// llama.cpp's llama-server, vLLM and SGLang accept (e.g.
+	// {"enable_thinking": false}). Scalar values are folded into Metadata, so
+	// they take the same path to the backend; an explicit Metadata key wins.
+	ChatTemplateKwargs map[string]any `json:"chat_template_kwargs,omitempty" yaml:"chat_template_kwargs"`
 }
 
 type ModelsDataResponse struct {
diff --git a/docs/content/advanced/model-configuration.md b/docs/content/advanced/model-configuration.md
index 62153f3438..98e92d7d9f 100644
--- a/docs/content/advanced/model-configuration.md
+++ b/docs/content/advanced/model-configuration.md
@@ -751,6 +751,19 @@ and (for `enable_thinking` / `reasoning_effort`) takes effect across every backe
 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.
 
+Clients written for llama.cpp's `llama-server`, vLLM or SGLang send the same switches as a
+top-level `chat_template_kwargs` object. LocalAI accepts that too and folds its scalar values
+into `metadata` (booleans become `"true"` / `"false"`, numbers become strings; nested objects
+and arrays are ignored). If a key appears in both, `metadata` wins:
+
+```json
+{
+  "model": "qwen3",
+  "messages": [{"role": "user", "content": "hi"}],
+  "chat_template_kwargs": { "enable_thinking": false }
+}
+```
+
 ### Multimodal Backend Options
 
 | Option | Type | Default | Description |
@@ -1134,8 +1147,9 @@ curl http://localhost:8080/v1/chat/completions \
 
 **Priority order:**
 1. Request-level `metadata.enable_thinking` (highest priority)
-2. Model config `reasoning.disable` (fallback)
-3. Auto-detected from model template (default)
+2. Request-level `chat_template_kwargs.enable_thinking` (the llama-server / vLLM form)
+3. Model config `reasoning.disable` (fallback)
+4. Auto-detected from model template (default)
 
 ## Pipeline Configuration