Add template_fill config type to generate model variants from one template

config-schema.json

@@ -235,6 +235,48 @@
                         "type": "boolean",
                         "default": false,
                         "description": "If true the model will not show up in /v1/models responses. It can still be used as normal in API requests."
+                    },
+                    "variants": {
+                        "type": "object",
+                        "additionalProperties": {
+                            "type": "object",
+                            "properties": {
+                                "cmdAdd": {
+                                    "type": "string",
+                                    "description": "Additional command line arguments to append or override. Arguments with the same flag name will override the base cmd."
+                                },
+                                "name": {
+                                    "type": "string",
+                                    "description": "Display name override for this variant."
+                                },
+                                "description": {
+                                    "type": "string",
+                                    "description": "Description override for this variant."
+                                },
+                                "env": {
+                                    "type": "array",
+                                    "items": {
+                                        "type": "string",
+                                        "pattern": "^[A-Z_][A-Z0-9_]*=.*$"
+                                    },
+                                    "description": "Additional environment variables for this variant."
+                                },
+                                "aliases": {
+                                    "type": "array",
+                                    "items": {
+                                        "type": "string",
+                                        "minLength": 1
+                                    },
+                                    "description": "Additional aliases for this variant."
+                                },
+                                "unlisted": {
+                                    "type": "boolean",
+                                    "description": "Override unlisted setting for this variant."
+                                }
+                            },
+                            "additionalProperties": false
+                        },
+                        "description": "Template-based configuration that generates multiple model variants. Each key becomes a suffix to the model name (e.g., 'model-variant'). Variant values can override cmd arguments and other settings. See issue #549."
                     }
                 }
             }

config.example.yaml

@@ -277,6 +277,51 @@ models:
     # - optional, default: undefined (use global setting)
     sendLoadingState: false
 
+  # Variants example (issue #549):
+  # Use variants to generate multiple model configurations from a single template.
+  # Each variant key becomes a suffix to the model name (e.g., "Qwen3.5-thinking_normal").
+  # Variant's cmdAdd can override or add command line arguments.
+  "Qwen3.5-35B-A3B":
+    cmd: |
+      llama-server
+      --port ${PORT}
+      --model /models/Qwen3.5-35B-A3B.gguf
+      --temp 0.8
+      --ctx-size 16384
+    name: "Qwen3.5 35B"
+    description: "Qwen3.5 thinking model with multiple parameter variants"
+
+    # variants: generates multiple model configurations from this template
+    # - each key becomes a model name suffix (model-variant)
+    # - cmdAdd: additional/override command line arguments
+    # - name, description, env, aliases, unlisted: override base settings
+    variants:
+      # Generates: Qwen3.5-35B-A3B-thinking_normal
+      thinking_normal:
+        cmdAdd: --temp 1.0
+        name: "Qwen3.5 35B (Thinking, Normal)"
+        description: "For general tasks with thinking enabled"
+
+      # Generates: Qwen3.5-35B-A3B-thinking_coding
+      thinking_coding:
+        cmdAdd: --temp 0.6
+        name: "Qwen3.5 35B (Thinking, Coding)"
+        description: "For coding tasks with thinking enabled"
+
+      # Generates: Qwen3.5-35B-A3B-nothinking_normal
+      nothinking_normal:
+        cmdAdd: "--temp 1.0 --chat-template-kwargs '{\"enable_thinking\": false}'"
+        name: "Qwen3.5 35B (No Thinking, Normal)"
+        description: "For general tasks without thinking"
+
+      # Generates: Qwen3.5-35B-A3B-nothinking_coding
+      nothinking_coding:
+        cmdAdd: "--temp 0.6 --chat-template-kwargs '{\"enable_thinking\": false}'"
+        name: "Qwen3.5 35B (No Thinking, Coding)"
+        description: "For coding tasks without thinking"
+        aliases:
+          - "qwen-code"
+
   # Unlisted model example:
   "qwen-unlisted":
     # unlisted: boolean, true or false

proxy/config/config.go

@@ -171,6 +171,38 @@ func (c *Config) FindConfig(modelName string) (ModelConfig, string, bool) {
 	}
 }
 
+// substituteTemplateRefsInConfig replaces template model IDs in groups.members and
+// hooks.on_startup.preload with the corresponding expanded variant IDs.
+func substituteTemplateRefsInConfig(c *Config, templateToVariants map[string][]string) {
+	if len(templateToVariants) == 0 {
+		return
+	}
+	for groupID := range c.Groups {
+		group := c.Groups[groupID]
+		var newMembers []string
+		for _, member := range group.Members {
+			if variantIDs, ok := templateToVariants[member]; ok {
+				newMembers = append(newMembers, variantIDs...)
+			} else {
+				newMembers = append(newMembers, member)
+			}
+		}
+		group.Members = newMembers
+		c.Groups[groupID] = group
+	}
+	if len(c.Hooks.OnStartup.Preload) > 0 {
+		var newPreload []string
+		for _, modelID := range c.Hooks.OnStartup.Preload {
+			if variantIDs, ok := templateToVariants[modelID]; ok {
+				newPreload = append(newPreload, variantIDs...)
+			} else {
+				newPreload = append(newPreload, modelID)
+			}
+		}
+		c.Hooks.OnStartup.Preload = newPreload
+	}
+}
+
 func LoadConfig(path string) (Config, error) {
 	file, err := os.Open(path)
 	if err != nil {
@@ -208,6 +240,18 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		return Config{}, err
 	}
 
+	// Expand model variants before any other processing
+	// This transforms template models with variants into individual model configs
+	if config.Models != nil {
+		expanded, err := ExpandVariants(config.Models)
+		if err != nil {
+			return Config{}, err
+		}
+		config.Models = expanded.Models
+		// Substitute template IDs in groups.members and hooks.on_startup.preload with variant IDs
+		substituteTemplateRefsInConfig(&config, expanded.TemplateToVariants)
+	}
+
 	if config.HealthCheckTimeout < 15 {
 		config.HealthCheckTimeout = 15
 	}

proxy/config/model_config.go

@@ -36,6 +36,33 @@ type ModelConfig struct {
 
 	// override global setting
 	SendLoadingState *bool `yaml:"sendLoadingState"`
+
+	// Variants: see issue #549
+	// Template-based configuration that generates multiple model variants
+	// Each variant key becomes a suffix to the model name (e.g., "model-variant")
+	// Variant values can override cmd arguments
+	Variants map[string]VariantConfig `yaml:"variants"`
+}
+
+// VariantConfig holds the configuration overrides for a model variant
+type VariantConfig struct {
+	// CmdAdd contains additional command line arguments to append or override
+	CmdAdd string `yaml:"cmdAdd"`
+
+	// Name override for this variant
+	Name string `yaml:"name"`
+
+	// Description override for this variant
+	Description string `yaml:"description"`
+
+	// Env additional environment variables for this variant
+	Env []string `yaml:"env"`
+
+	// Aliases additional aliases for this variant
+	Aliases []string `yaml:"aliases"`
+
+	// Unlisted override for this variant
+	Unlisted *bool `yaml:"unlisted"`
 }
 
 func (m *ModelConfig) UnmarshalYAML(unmarshal func(interface{}) error) error {