Model capabilities 734 (#842)

internal/config,server: implement model capabilities

- define the capabilities of a model using a simple config block on the
model
- v1/models renders out capabilities to be compatible with openrouter,
huggingface chat, and mistral formats for broader compatibility
- add support for capabilities in UI

Fixes #734

config-schema.json

@@ -378,6 +378,59 @@
                     },
                     "timeouts": {
                         "$ref": "#/definitions/timeouts"
+                    },
+                    "capabilities": {
+                        "type": "object",
+                        "properties": {
+                            "in": {
+                                "type": "array",
+                                "minItems": 1,
+                                "uniqueItems": true,
+                                "default": [],
+                                "items": {
+                                    "type": "string",
+                                    "enum": [
+                                        "text",
+                                        "audio",
+                                        "image"
+                                    ]
+                                },
+                                "description": "List of input modalities understood by the model."
+                            },
+                            "out": {
+                                "type": "array",
+                                "minItems": 1,
+                                "uniqueItems": true,
+                                "default": [],
+                                "items": {
+                                    "type": "string",
+                                    "enum": [
+                                        "text",
+                                        "audio",
+                                        "image"
+                                    ]
+                                },
+                                "description": "List of output modalities generated by the model."
+                            },
+                            "tools": {
+                                "type": "boolean",
+                                "default": false,
+                                "description": "Whether the model supports function calling."
+                            },
+                            "reranker": {
+                                "type": "boolean",
+                                "default": false,
+                                "description": "Whether the model supports the /v1/rerank endpoint."
+                            },
+                            "context": {
+                                "type": "integer",
+                                "minimum": 0,
+                                "default": 0,
+                                "description": "Maximum token context length supported by the model."
+                            }
+                        },
+                        "additionalProperties": false,
+                        "description": "Defines what the model accepts for input, output and other metadata. Used in v1/models to inform clients what the model can do. An empty capabilities block (all zero values) is treated as not configured."
                     }
                 }
             }

config.example.yaml

@@ -312,6 +312,37 @@ models:
       tlsHandshake: 10
       idleConn: 90
 
+    # capabilities: defines what the model accepts for input, output and other metadata
+    # - optional; omitted or all-zero means no capabilities
+    # - used in v1/models to inform clients what the model can do
+    capabilities:
+      # in: list of modalities understood by the model
+      # - default: []
+      # - valid: text, audio, image
+      in:
+        - text
+        - audio
+        - image
+      # out: list of modalities generated by the model
+      # - default: []
+      # - valid: text, audio, image
+      out:
+        - text
+        - audio
+        - image
+      # tools: the model supports function calling
+      # - default: false
+      tools: true
+
+      # reranker: the model supports the /v1/rerank endpoint
+      # - default: false
+      reranker: false
+
+      # context: the maximum token context length supported
+      # - default: 0
+      # - must be an integer > 0
+      context: 32000
+
   # Unlisted model example:
   "qwen-unlisted":
     # unlisted: boolean, true or false

internal/config/config.go

@@ -447,6 +447,10 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			}
 		}
 
+		if err = modelConfig.Capabilities.Validate(); err != nil {
+			return Config{}, fmt.Errorf("model %s: %w", modelId, err)
+		}
+
 		// Validate SetParamsByID keys and values
 		for key, paramMap := range modelConfig.Filters.SetParamsByID {
 			if matches := macroPatternRegex.FindAllStringSubmatch(key, -1); len(matches) > 0 {

internal/config/model_config.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"errors"
+	"fmt"
 	"runtime"
 )
 
@@ -9,6 +10,47 @@ const (
 	MODEL_CONFIG_DEFAULT_TTL = -1
 )
 
+var validModalities = map[string]struct{}{
+	"text":  {},
+	"audio": {},
+	"image": {},
+}
+
+// ModelCapConfig defines what modalities and features a model supports.
+// Used in /v1/models to inform clients. An empty block (all zero values) is
+// treated as not configured.
+type ModelCapConfig struct {
+	In       []string `yaml:"in"`
+	Out      []string `yaml:"out"`
+	Tools    bool     `yaml:"tools"`
+	Reranker bool     `yaml:"reranker"`
+	Context  int      `yaml:"context"`
+}
+
+// Empty returns true when all fields are at their zero values.
+func (c ModelCapConfig) Empty() bool {
+	return len(c.In) == 0 && len(c.Out) == 0 && !c.Tools && !c.Reranker && c.Context == 0
+}
+
+// Validate checks that all modality values are recognized and context is
+// non-negative. Returns an error if any value is invalid.
+func (c ModelCapConfig) Validate() error {
+	for _, m := range c.In {
+		if _, ok := validModalities[m]; !ok {
+			return fmt.Errorf("capabilities.in: invalid modality %q, must be one of: text, audio, image", m)
+		}
+	}
+	for _, m := range c.Out {
+		if _, ok := validModalities[m]; !ok {
+			return fmt.Errorf("capabilities.out: invalid modality %q, must be one of: text, audio, image", m)
+		}
+	}
+	if c.Context < 0 {
+		return errors.New("capabilities.context: must be >= 0")
+	}
+	return nil
+}
+
 // TimeoutsConfig holds timeout settings for proxy connections
 // 0 = no timeout
 type TimeoutsConfig struct {
@@ -55,6 +97,9 @@ type ModelConfig struct {
 	// Timeout settings for proxy connections
 	Timeouts TimeoutsConfig `yaml:"timeouts"`
 
+	// Capabilities defines what modalities and features the model supports.
+	Capabilities ModelCapConfig `yaml:"capabilities"`
+
 	// Copy of HealthCheckTimeout from global config
 	HealthCheckTimeout int `yaml:"healthCheckTimeout"`
 }

internal/config/model_config_test.go

@@ -170,3 +170,167 @@ models:
 	assert.Equal(t, 0.7, setParams["temperature"])
 	assert.Equal(t, 0.9, setParams["top_p"])
 }
+
+func TestConfig_ModelCapabilities(t *testing.T) {
+	t.Run("all fields", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    capabilities:
+      in:
+        - text
+        - audio
+        - image
+      out:
+        - text
+        - audio
+        - image
+      tools: true
+      context: 32000
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+
+		mc := config.Models["model1"]
+		assert.False(t, mc.Capabilities.Empty())
+		assert.Equal(t, []string{"text", "audio", "image"}, mc.Capabilities.In)
+		assert.Equal(t, []string{"text", "audio", "image"}, mc.Capabilities.Out)
+		assert.True(t, mc.Capabilities.Tools)
+		assert.Equal(t, 32000, mc.Capabilities.Context)
+	})
+
+	t.Run("partial fields", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    capabilities:
+      tools: true
+      context: 8192
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+
+		mc := config.Models["model1"]
+		assert.False(t, mc.Capabilities.Empty())
+		assert.Nil(t, mc.Capabilities.In)
+		assert.Nil(t, mc.Capabilities.Out)
+		assert.True(t, mc.Capabilities.Tools)
+		assert.Equal(t, 8192, mc.Capabilities.Context)
+	})
+
+	t.Run("not set", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+
+		mc := config.Models["model1"]
+		assert.True(t, mc.Capabilities.Empty())
+	})
+
+	t.Run("tools false is empty", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    capabilities:
+      tools: false
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+
+		mc := config.Models["model1"]
+		assert.True(t, mc.Capabilities.Empty())
+	})
+
+	t.Run("reranker true is not empty", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    capabilities:
+      reranker: true
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+
+		mc := config.Models["model1"]
+		assert.False(t, mc.Capabilities.Empty())
+		assert.True(t, mc.Capabilities.Reranker)
+	})
+
+	t.Run("reranker false is empty", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    capabilities:
+      reranker: false
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+
+		mc := config.Models["model1"]
+		assert.True(t, mc.Capabilities.Empty())
+	})
+}
+
+func TestConfig_ModelCapabilities_Validate(t *testing.T) {
+	t.Run("valid_modalities", func(t *testing.T) {
+		caps := ModelCapConfig{
+			In:      []string{"text", "image"},
+			Out:     []string{"text", "audio"},
+			Tools:   true,
+			Context: 100000,
+		}
+		assert.NoError(t, caps.Validate())
+	})
+
+	t.Run("empty_is_valid", func(t *testing.T) {
+		caps := ModelCapConfig{}
+		assert.NoError(t, caps.Validate())
+	})
+
+	t.Run("invalid_in_modality", func(t *testing.T) {
+		caps := ModelCapConfig{In: []string{"video"}}
+		err := caps.Validate()
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "capabilities.in")
+		assert.Contains(t, err.Error(), "video")
+	})
+
+	t.Run("invalid_out_modality", func(t *testing.T) {
+		caps := ModelCapConfig{Out: []string{"video"}}
+		err := caps.Validate()
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "capabilities.out")
+		assert.Contains(t, err.Error(), "video")
+	})
+
+	t.Run("negative_context", func(t *testing.T) {
+		caps := ModelCapConfig{Context: -1}
+		err := caps.Validate()
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "capabilities.context")
+	})
+
+	t.Run("rejects_invalid_at_load", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    capabilities:
+      in:
+        - text
+        - video
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "video")
+	})
+}

internal/server/api.go

@@ -23,16 +23,121 @@ type modelRecord struct {
 	OwnedBy             string         `json:"owned_by"`
 	Name                string         `json:"name,omitempty"`
 	Description         string         `json:"description,omitempty"`
+	Architecture        map[string]any `json:"architecture,omitempty"`
+	Capabilities        map[string]any `json:"capabilities,omitempty"`
+	SupportedParameters []string       `json:"supported_parameters,omitempty"`
+	ContextLength       int            `json:"context_length,omitempty"`
 	Meta                map[string]any `json:"meta,omitempty"`
 }
 
+// cappedMetadataKeys are top-level /v1/models fields produced by the
+// capabilities renderer. If a model's metadata block defines any of these
+// keys, the renderer's values win and the metadata keys are dropped.
+var cappedMetadataKeys = map[string]struct{}{
+	"architecture":         {},
+	"capabilities":         {},
+	"supported_parameters": {},
+	"context_length":       {},
+}
+
+// renderCapabilities converts a model's capabilities config into additional
+// /v1/models fields. Returns zero values when caps.Empty() is true.
+func renderCapabilities(caps config.ModelCapConfig) (arch map[string]any, capsMap map[string]any, params []string, ctxLen int) {
+	if caps.Empty() {
+		return
+	}
+
+	hasIn := len(caps.In) > 0
+	hasOut := len(caps.Out) > 0
+
+	if hasIn || hasOut {
+		arch = make(map[string]any)
+	}
+	if hasIn {
+		arch["input_modalities"] = caps.In
+	}
+	if hasOut {
+		arch["output_modalities"] = caps.Out
+	}
+	if hasIn && hasOut {
+		arch["modality"] = strings.Join(caps.In, "+") + "->" + strings.Join(caps.Out, "+")
+	}
+
+	// Build capabilities map only if there's something to put in it.
+	if hasIn || hasOut || caps.Tools || caps.Reranker {
+		capsMap = make(map[string]any)
+	}
+
+	if hasIn {
+		if contains(caps.In, "image") {
+			capsMap["vision"] = true
+		}
+	}
+	if hasIn && hasOut {
+		if contains(caps.In, "audio") && contains(caps.Out, "text") {
+			capsMap["audio_transcriptions"] = true
+		}
+		if contains(caps.In, "text") && contains(caps.Out, "audio") {
+			capsMap["audio_speech"] = true
+		}
+		if contains(caps.In, "text") && contains(caps.Out, "image") {
+			capsMap["image_generation"] = true
+		}
+		if contains(caps.In, "image") && contains(caps.Out, "image") {
+			capsMap["image_to_image"] = true
+		}
+	}
+
+	if caps.Tools {
+		capsMap["function_calling"] = true
+		params = []string{"tools", "tool_choice"}
+	}
+
+	if caps.Reranker {
+		capsMap["reranker"] = true
+	}
+
+	if caps.Context > 0 {
+		ctxLen = caps.Context
+	}
+
+	return
+}
+
+// contains reports whether s is present in ss.
+func contains(ss []string, s string) bool {
+	for _, v := range ss {
+		if v == s {
+			return true
+		}
+	}
+	return false
+}
+
+// filterCappedMetadata returns metadata with renderer-owned keys removed.
+func filterCappedMetadata(md map[string]any) map[string]any {
+	if len(md) == 0 {
+		return nil
+	}
+	filtered := make(map[string]any, len(md))
+	for k, v := range md {
+		if _, capped := cappedMetadataKeys[k]; !capped {
+			filtered[k] = v
+		}
+	}
+	if len(filtered) == 0 {
+		return nil
+	}
+	return filtered
+}
+
 // handleListModels serves the OpenAI-compatible model listing: local models
 // (with optional aliases) plus peer models.
 func (s *Server) handleListModels(w http.ResponseWriter, r *http.Request) {
 	created := time.Now().Unix()
 	data := make([]modelRecord, 0, len(s.cfg.Models))
 
-	newRecord := func(id, name, description string, metadata map[string]any) modelRecord {
+	newRecord := func(id, name, description string, metadata map[string]any, caps config.ModelCapConfig) modelRecord {
 		rec := modelRecord{
 			ID:          id,
 			Object:      "model",
@@ -41,6 +146,10 @@ func (s *Server) handleListModels(w http.ResponseWriter, r *http.Request) {
 			Name:        strings.TrimSpace(name),
 			Description: strings.TrimSpace(description),
 		}
+		rec.Architecture, rec.Capabilities, rec.SupportedParameters, rec.ContextLength = renderCapabilities(caps)
+		if !caps.Empty() {
+			metadata = filterCappedMetadata(metadata)
+		}
 		if len(metadata) > 0 {
 			rec.Meta = map[string]any{"llamaswap": metadata}
 		}
@@ -51,12 +160,12 @@ func (s *Server) handleListModels(w http.ResponseWriter, r *http.Request) {
 		if mc.Unlisted {
 			continue
 		}
-		data = append(data, newRecord(id, mc.Name, mc.Description, mc.Metadata))
+		data = append(data, newRecord(id, mc.Name, mc.Description, mc.Metadata, mc.Capabilities))
 
 		if s.cfg.IncludeAliasesInList {
 			for _, alias := range mc.Aliases {
 				if alias := strings.TrimSpace(alias); alias != "" {
-					data = append(data, newRecord(alias, mc.Name, mc.Description, mc.Metadata))
+					data = append(data, newRecord(alias, mc.Name, mc.Description, mc.Metadata, mc.Capabilities))
 				}
 			}
 		}
@@ -64,7 +173,7 @@ func (s *Server) handleListModels(w http.ResponseWriter, r *http.Request) {
 
 	for peerID, peer := range s.cfg.Peers {
 		for _, modelID := range peer.Models {
-			data = append(data, newRecord(modelID, peerID+": "+modelID, "", map[string]any{"peerID": peerID}))
+			data = append(data, newRecord(modelID, peerID+": "+modelID, "", map[string]any{"peerID": peerID}, config.ModelCapConfig{}))
 		}
 	}
 

internal/server/api_test.go

@@ -157,3 +157,262 @@ func TestServer_Redirects(t *testing.T) {
 		}
 	}
 }
+
+func TestServer_HandleListModels_Capabilities(t *testing.T) {
+	newServer := func(mc config.ModelConfig) *Server {
+		s := newTestServer(newStubRouter(nil, ""), newStubRouter(nil, ""))
+		s.cfg = config.Config{Models: map[string]config.ModelConfig{"m": mc}}
+		return s
+	}
+	getModel := func(t *testing.T, s *Server) modelRecord {
+		t.Helper()
+		w := httptest.NewRecorder()
+		s.ServeHTTP(w, httptest.NewRequest(http.MethodGet, "/v1/models", nil))
+		var resp struct {
+			Data []modelRecord `json:"data"`
+		}
+		if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+			t.Fatalf("decode: %v", err)
+		}
+		if len(resp.Data) != 1 {
+			t.Fatalf("expected 1 model, got %d", len(resp.Data))
+		}
+		return resp.Data[0]
+	}
+
+	t.Run("all_fields", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{
+				In:      []string{"text", "image"},
+				Out:     []string{"text", "audio"},
+				Tools:   true,
+				Context: 100000,
+			},
+		}))
+		if m.Architecture == nil {
+			t.Fatal("architecture is nil")
+		}
+		if !anySliceStrEqual(m.Architecture["input_modalities"], []string{"text", "image"}) {
+			t.Errorf("input_modalities = %v", m.Architecture["input_modalities"])
+		}
+		if !anySliceStrEqual(m.Architecture["output_modalities"], []string{"text", "audio"}) {
+			t.Errorf("output_modalities = %v", m.Architecture["output_modalities"])
+		}
+		if m.Architecture["modality"] != "text+image->text+audio" {
+			t.Errorf("modality = %v", m.Architecture["modality"])
+		}
+		if m.Capabilities == nil || m.Capabilities["vision"] != true {
+			t.Errorf("vision = %v", m.Capabilities)
+		}
+		if m.Capabilities["audio_speech"] != true {
+			t.Errorf("audio_speech = %v", m.Capabilities["audio_speech"])
+		}
+		if m.Capabilities["function_calling"] != true {
+			t.Errorf("function_calling = %v", m.Capabilities["function_calling"])
+		}
+		if !stringSliceEqual(m.SupportedParameters, []string{"tools", "tool_choice"}) {
+			t.Errorf("supported_parameters = %v", m.SupportedParameters)
+		}
+		if m.ContextLength != 100000 {
+			t.Errorf("context_length = %d", m.ContextLength)
+		}
+	})
+
+	t.Run("in_only", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{In: []string{"text", "image"}},
+		}))
+		if m.Architecture == nil {
+			t.Fatal("architecture is nil")
+		}
+		if _, ok := m.Architecture["output_modalities"]; ok {
+			t.Error("should not have output_modalities")
+		}
+		if _, ok := m.Architecture["modality"]; ok {
+			t.Error("should not have modality")
+		}
+		if m.Capabilities == nil || m.Capabilities["vision"] != true {
+			t.Error("expected vision: true")
+		}
+		if m.SupportedParameters != nil {
+			t.Error("should not have supported_parameters")
+		}
+		if m.ContextLength != 0 {
+			t.Error("should not have context_length")
+		}
+	})
+
+	t.Run("out_only", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{Out: []string{"audio"}},
+		}))
+		if m.Architecture == nil {
+			t.Fatal("architecture is nil")
+		}
+		if _, ok := m.Architecture["input_modalities"]; ok {
+			t.Error("should not have input_modalities")
+		}
+		if len(m.Capabilities) > 0 {
+			t.Errorf("expected no capabilities, got %v", m.Capabilities)
+		}
+	})
+
+	t.Run("tools", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{Tools: true},
+		}))
+		if m.Capabilities == nil || m.Capabilities["function_calling"] != true {
+			t.Error("expected function_calling: true")
+		}
+		if !stringSliceEqual(m.SupportedParameters, []string{"tools", "tool_choice"}) {
+			t.Errorf("supported_parameters = %v", m.SupportedParameters)
+		}
+		if m.Architecture != nil {
+			t.Error("should not have architecture")
+		}
+	})
+
+	t.Run("reranker", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{Reranker: true},
+		}))
+		if m.Capabilities == nil || m.Capabilities["reranker"] != true {
+			t.Error("expected reranker: true")
+		}
+		if m.Architecture != nil {
+			t.Error("should not have architecture")
+		}
+	})
+
+	t.Run("context", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{Context: 32768},
+		}))
+		if m.ContextLength != 32768 {
+			t.Errorf("context_length = %d", m.ContextLength)
+		}
+		if m.Architecture != nil {
+			t.Error("should not have architecture")
+		}
+	})
+
+	t.Run("audio_transcriptions", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{In: []string{"audio"}, Out: []string{"text"}},
+		}))
+		if m.Capabilities == nil || m.Capabilities["audio_transcriptions"] != true {
+			t.Error("expected audio_transcriptions: true")
+		}
+	})
+
+	t.Run("image_generation", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{In: []string{"text"}, Out: []string{"image"}},
+		}))
+		if m.Capabilities == nil || m.Capabilities["image_generation"] != true {
+			t.Error("expected image_generation: true")
+		}
+	})
+
+	t.Run("image_to_image", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{In: []string{"image"}, Out: []string{"image"}},
+		}))
+		if m.Capabilities == nil || m.Capabilities["image_to_image"] != true {
+			t.Error("expected image_to_image: true")
+		}
+	})
+
+	t.Run("empty_skip", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{}))
+		if m.Architecture != nil {
+			t.Error("should not have architecture")
+		}
+		if m.Capabilities != nil {
+			t.Error("should not have capabilities")
+		}
+		if m.SupportedParameters != nil {
+			t.Error("should not have supported_parameters")
+		}
+		if m.ContextLength != 0 {
+			t.Error("should not have context_length")
+		}
+	})
+
+	t.Run("metadata_precedence", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Capabilities: config.ModelCapConfig{In: []string{"text"}},
+			Metadata: map[string]any{
+				"architecture":   "should-be-dropped",
+				"custom_field":   "should-remain",
+				"capabilities":   "also-dropped",
+				"other_metadata": "also-remain",
+			},
+		}))
+		if m.Architecture == nil || m.Architecture["input_modalities"] == nil {
+			t.Fatal("architecture should be rendered, not from metadata")
+		}
+		if m.Meta == nil || m.Meta["llamaswap"] == nil {
+			t.Fatal("meta.llamaswap should exist")
+		}
+		meta := m.Meta["llamaswap"].(map[string]any)
+		if _, ok := meta["architecture"]; ok {
+			t.Error("architecture should be filtered from metadata")
+		}
+		if _, ok := meta["custom_field"]; !ok {
+			t.Error("custom_field should remain in metadata")
+		}
+	})
+
+	t.Run("metadata_passthrough_no_caps", func(t *testing.T) {
+		m := getModel(t, newServer(config.ModelConfig{
+			Metadata: map[string]any{
+				"architecture":   "preserved",
+				"context_length": 4096,
+				"capabilities":   "preserved",
+				"custom_field":   "preserved",
+			},
+		}))
+		if m.Architecture != nil {
+			t.Error("should not have architecture when caps is empty")
+		}
+		if m.Meta == nil || m.Meta["llamaswap"] == nil {
+			t.Fatal("meta.llamaswap should exist")
+		}
+		meta := m.Meta["llamaswap"].(map[string]any)
+		if _, ok := meta["architecture"]; !ok {
+			t.Error("architecture should be preserved in metadata when caps is empty")
+		}
+		if _, ok := meta["context_length"]; !ok {
+			t.Error("context_length should be preserved in metadata when caps is empty")
+		}
+	})
+}
+
+func stringSliceEqual(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+func anySliceStrEqual(v any, want []string) bool {
+	arr, ok := v.([]any)
+	if !ok {
+		return false
+	}
+	if len(arr) != len(want) {
+		return false
+	}
+	for i := range arr {
+		if s, ok := arr[i].(string); !ok || s != want[i] {
+			return false
+		}
+	}
+	return true
+}

internal/server/apigroup.go

@@ -24,6 +24,7 @@ type apiModel struct {
 	Unlisted     bool           `json:"unlisted"`
 	PeerID       string         `json:"peerID"`
 	Aliases      []string       `json:"aliases,omitempty"`
+	Capabilities map[string]any `json:"capabilities,omitempty"`
 }
 
 // modelStatus returns every configured model joined with its current process
@@ -44,6 +45,7 @@ func (s *Server) modelStatus() []apiModel {
 		if st, ok := running[id]; ok {
 			state = string(st)
 		}
+		_, capsMap, _, _ := renderCapabilities(mc.Capabilities)
 		models = append(models, apiModel{
 			Id:           id,
 			Name:         mc.Name,
@@ -51,6 +53,7 @@ func (s *Server) modelStatus() []apiModel {
 			State:        state,
 			Unlisted:     mc.Unlisted,
 			Aliases:      mc.Aliases,
+			Capabilities: capsMap,
 		})
 	}
 

ui-svelte/src/components/playground/AudioInterface.svelte

@@ -145,7 +145,7 @@
 <div class="flex flex-col h-full">
   <!-- Model selector -->
   <div class="shrink-0 flex flex-wrap gap-2 mb-4">
-    <ModelSelector bind:value={$selectedModelStore} placeholder="Select an audio model..." disabled={isTranscribing} />
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select an audio model..." disabled={isTranscribing} capabilities={["audio_transcriptions"]} />
   </div>
 
   <!-- Empty state for no models configured -->

ui-svelte/src/components/playground/ImageInterface.svelte

@@ -193,7 +193,7 @@
 <div class="flex flex-col h-full">
   <!-- Model selector and mode toggle -->
   <div class="shrink-0 flex flex-wrap gap-2 mb-4">
-    <ModelSelector bind:value={$selectedModelStore} placeholder="Select an image model..." disabled={isGenerating} />
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select an image model..." disabled={isGenerating} capabilities={["image_generation", "image_to_image"]} matchAny={true} />
 
     <select
       class="px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"

ui-svelte/src/components/playground/ModelSelector.svelte

@@ -6,12 +6,15 @@
     value: string;
     placeholder?: string;
     disabled?: boolean;
+    capabilities?: string[];
+    matchAny?: boolean;
   }
 
-  let { value = $bindable(), placeholder = "Select a model...", disabled = false }: Props = $props();
+  let { value = $bindable(), placeholder = "Select a model...", disabled = false, capabilities, matchAny = false }: Props = $props();
 
-  let grouped = $derived(groupModels($models));
-  let hasModels = $derived(grouped.local.length > 0 || Object.keys(grouped.peersByProvider).length > 0);
+  let grouped = $derived(groupModels($models, capabilities, matchAny));
+  let hasMatching = $derived(grouped.localMatching.length > 0);
+  let hasModels = $derived(hasMatching || grouped.local.length > 0 || Object.keys(grouped.peersByProvider).length > 0);
 </script>
 
 {#if hasModels}
@@ -21,6 +24,18 @@
     {disabled}
   >
     <option value="">{placeholder}</option>
+    {#if hasMatching}
+      <optgroup label="Matching Capabilities">
+        {#each grouped.localMatching as model (model.id)}
+          <option value={model.id}>{model.id}</option>
+          {#if model.aliases}
+            {#each model.aliases as alias (alias)}
+              <option value={alias}>  ↳ {alias}</option>
+            {/each}
+          {/if}
+        {/each}
+      </optgroup>
+    {/if}
     {#if grouped.local.length > 0}
       <optgroup label="Local">
         {#each grouped.local as model (model.id)}

ui-svelte/src/components/playground/RerankInterface.svelte

@@ -264,7 +264,7 @@
 <div class="flex flex-col h-full">
   <!-- Top bar: model selector + query input (table mode) + mode toggle -->
   <div class="shrink-0 flex flex-wrap gap-2 mb-4">
-    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a rerank model..." disabled={isLoading} />
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a rerank model..." disabled={isLoading} capabilities={["reranker"]} />
     {#if editorMode === "table"}
       <input
         type="text"

ui-svelte/src/components/playground/SpeechInterface.svelte

@@ -206,7 +206,7 @@
 <div class="flex flex-col h-full">
   <!-- Model and voice selectors -->
   <div class="shrink-0 flex gap-2 mb-4">
-    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a speech model..." disabled={isGenerating} />
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a speech model..." disabled={isGenerating} capabilities={["audio_speech"]} />
     <div class="flex gap-2">
       <select
         class="shrink-0 px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"

ui-svelte/src/lib/modelUtils.test.ts

@@ -0,0 +1,113 @@
+import { describe, it, expect } from "vitest";
+import { matchesCapabilities, groupModels } from "./modelUtils";
+import type { Model } from "./types";
+
+function makeModel(overrides: Partial<Model> = {}): Model {
+  return {
+    id: "test-model",
+    state: "ready",
+    name: "Test Model",
+    description: "",
+    unlisted: false,
+    peerID: "",
+    ...overrides,
+  };
+}
+
+describe("matchesCapabilities", () => {
+  it("returns true when required is empty", () => {
+    const model = makeModel();
+    expect(matchesCapabilities(model, [])).toBe(true);
+  });
+
+  it("returns false when model has no capabilities", () => {
+    const model = makeModel();
+    expect(matchesCapabilities(model, ["vision"])).toBe(false);
+  });
+
+  it("returns false when model has empty capabilities object", () => {
+    const model = makeModel({ capabilities: {} });
+    expect(matchesCapabilities(model, ["vision"])).toBe(false);
+  });
+
+  it("returns true when model has the single required capability", () => {
+    const model = makeModel({ capabilities: { vision: true } });
+    expect(matchesCapabilities(model, ["vision"])).toBe(true);
+  });
+
+  it("returns false when model lacks the required capability", () => {
+    const model = makeModel({ capabilities: { vision: true } });
+    expect(matchesCapabilities(model, ["audio_transcriptions"])).toBe(false);
+  });
+
+  it("AND semantics: returns true only when all required are present", () => {
+    const model = makeModel({ capabilities: { vision: true, audio_transcriptions: true } });
+    expect(matchesCapabilities(model, ["vision", "audio_transcriptions"])).toBe(true);
+    expect(matchesCapabilities(model, ["vision", "reranker"])).toBe(false);
+  });
+
+  it("matchAny=true: returns true when at least one required is present", () => {
+    const model = makeModel({ capabilities: { vision: true } });
+    expect(matchesCapabilities(model, ["vision", "reranker"], true)).toBe(true);
+    expect(matchesCapabilities(model, ["audio_transcriptions", "reranker"], true)).toBe(false);
+  });
+
+  it("matchAny=true with empty required returns true", () => {
+    const model = makeModel();
+    expect(matchesCapabilities(model, [], true)).toBe(true);
+  });
+});
+
+describe("groupModels", () => {
+  const models: Model[] = [
+    makeModel({ id: "chat-model", capabilities: { vision: true } }),
+    makeModel({ id: "audio-model", capabilities: { audio_transcriptions: true } }),
+    makeModel({ id: "no-caps-model" }),
+    makeModel({ id: "peer-model", peerID: "peer1" }),
+    makeModel({ id: "unlisted-model", unlisted: true, capabilities: { vision: true } }),
+  ];
+
+  it("filters out unlisted models", () => {
+    const result = groupModels(models);
+    expect(result.localMatching.length + result.local.length).toBe(3);
+    expect([...result.localMatching, ...result.local].every((m) => !m.unlisted)).toBe(true);
+  });
+
+  it("separates peer models into peersByProvider", () => {
+    const result = groupModels(models);
+    expect(result.peersByProvider["peer1"]).toHaveLength(1);
+    expect(result.peersByProvider["peer1"][0].id).toBe("peer-model");
+  });
+
+  it("without capabilities, all local models go to local (non-matching)", () => {
+    const result = groupModels(models);
+    expect(result.localMatching).toHaveLength(0);
+    expect(result.local).toHaveLength(3);
+  });
+
+  it("with capabilities, matching models go to localMatching", () => {
+    const result = groupModels(models, ["vision"]);
+    expect(result.localMatching).toHaveLength(1);
+    expect(result.localMatching[0].id).toBe("chat-model");
+    expect(result.local).toHaveLength(2);
+  });
+
+  it("with capabilities, models without capabilities go to local", () => {
+    const result = groupModels(models, ["vision"]);
+    expect(result.local.find((m) => m.id === "no-caps-model")).toBeDefined();
+  });
+
+  it("with matchAny, matches models with any listed capability", () => {
+    const result = groupModels(models, ["vision", "audio_transcriptions"], true);
+    expect(result.localMatching).toHaveLength(2);
+    expect(result.localMatching.map((m) => m.id)).toContain("chat-model");
+    expect(result.localMatching.map((m) => m.id)).toContain("audio-model");
+    expect(result.local).toHaveLength(1);
+  });
+
+  it("with empty capabilities array, all local go to local (non-matching)", () => {
+    const result = groupModels(models, []);
+    expect(result.localMatching).toHaveLength(0);
+    expect(result.local).toHaveLength(3);
+  });
+});

ui-svelte/src/lib/modelUtils.ts

@@ -2,14 +2,40 @@ import type { Model } from "./types";
 
 export interface GroupedModels {
   local: Model[];
+  localMatching: Model[];
   peersByProvider: Record<string, Model[]>;
 }
 
-export function groupModels(models: Model[]): GroupedModels {
+export function matchesCapabilities(model: Model, required: string[], matchAny = false): boolean {
+  if (!required.length) return true;
+  if (!model.capabilities) return false;
+  const caps = model.capabilities as Record<string, boolean>;
+  if (matchAny) {
+    return required.some((cap) => caps[cap] === true);
+  }
+  return required.every((cap) => caps[cap] === true);
+}
+
+export function groupModels(models: Model[], capabilities?: string[], matchAny = false): GroupedModels {
   const available = models.filter((m) => !m.unlisted);
   const local = available.filter((m) => !m.peerID);
   const peerModels = available.filter((m) => m.peerID);
 
+  let localMatching: Model[] = [];
+  let localRest: Model[] = [];
+
+  if (capabilities && capabilities.length > 0) {
+    for (const model of local) {
+      if (matchesCapabilities(model, capabilities, matchAny)) {
+        localMatching.push(model);
+      } else {
+        localRest.push(model);
+      }
+    }
+  } else {
+    localRest = local;
+  }
+
   const peersByProvider = peerModels.reduce(
     (acc, model) => {
       const peerId = model.peerID || "unknown";
@@ -20,5 +46,5 @@ export function groupModels(models: Model[]): GroupedModels {
     {} as Record<string, Model[]>
   );
 
-  return { local, peersByProvider };
+  return { local: localRest, localMatching, peersByProvider };
 }

ui-svelte/src/lib/types.ts

@@ -2,6 +2,16 @@ export type ConnectionState = "connected" | "connecting" | "disconnected";
 
 export type ModelStatus = "ready" | "starting" | "stopping" | "stopped" | "shutdown" | "unknown";
 
+export interface ModelCapabilities {
+  vision?: boolean;
+  audio_transcriptions?: boolean;
+  audio_speech?: boolean;
+  image_generation?: boolean;
+  image_to_image?: boolean;
+  function_calling?: boolean;
+  reranker?: boolean;
+}
+
 export interface Model {
   id: string;
   state: ModelStatus;
@@ -10,6 +20,7 @@ export interface Model {
   unlisted: boolean;
   peerID: string;
   aliases?: string[];
+  capabilities?: ModelCapabilities;
 }
 
 export interface TokenMetrics {