Letta · Schema
LLMConfig
Configuration for Language Model (LLM) connection and generation parameters. .. deprecated:: LLMConfig is deprecated and should not be used as an input or return type in API calls. Use the schemas in letta.schemas.model (ModelSettings, OpenAIModelSettings, etc.) instead. For conversion, use the _to_model() method or Model._from_llm_config() method.
AIAgentsStateful AgentsMemoryMemGPTContinual LearningMCPMulti-AgentRAGOpen Source
Properties
| Name | Type | Description |
|---|---|---|
| model | string | LLM model name. |
| display_name | object | A human-friendly display name for the model. |
| model_endpoint_type | string | The endpoint type for the model. |
| model_endpoint | object | The endpoint for the model. |
| provider_name | object | The provider name for the model. |
| provider_category | object | The provider category for the model. |
| model_wrapper | object | The wrapper for the model. |
| context_window | integer | The context window size for the model. |
| put_inner_thoughts_in_kwargs | object | Puts 'inner_thoughts' as a kwarg in the function call if this is set to True. This helps with function calling performance and also the generation of inner thoughts. |
| handle | object | The handle for this config, in the format provider/model-name. |
| temperature | number | The temperature to use when generating text with the model. A higher temperature will result in more random text. |
| max_tokens | object | The maximum number of tokens to generate. If not set, the model will use its default value. |
| enable_reasoner | boolean | Whether or not the model should use extended thinking if it is a 'reasoning' style model |
| reasoning_effort | object | The reasoning effort to use when generating text reasoning models |
| max_reasoning_tokens | integer | Configurable thinking budget for extended thinking. Used for enable_reasoner and also for Google Vertex models like Gemini 2.5 Flash. Minimum value is 1024 when used with enable_reasoner. |
| effort | object | The effort level for Anthropic models that support it (Opus 4.5, Opus 4.6). Controls token spending and thinking behavior. Not setting this gives similar performance to 'high'. |
| frequency_penalty | object | Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim. From OpenAI: Number between -2.0 and 2.0. |
| compatibility_type | object | The framework compatibility type for the model. |
| verbosity | object | Soft control for how verbose model output should be, used for GPT-5 models. |
| tier | object | The cost tier for the model (cloud only). |
| parallel_tool_calls | object | Deprecated: Use model_settings to configure parallel tool calls instead. If set to True, enables parallel tool calling. Defaults to False. |
| response_format | object | The response format for the model's output. Supports text, json_object, and json_schema (structured outputs). Can be set via model_settings. |
| strict | boolean | Enable strict mode for tool calling. When true, tool schemas include strict: true and additionalProperties: false, guaranteeing tool outputs match JSON schemas. |
| return_logprobs | boolean | Whether to return log probabilities of the output tokens. Useful for RL training. |
| top_logprobs | object | Number of most likely tokens to return at each position (0-20). Requires return_logprobs=True. |
| return_token_ids | boolean | Whether to return token IDs for all LLM generations via SGLang native endpoint. Required for multi-turn RL training with loss masking. Only works with SGLang provider. |
| tool_call_parser | object | SGLang tool call parser name (e.g. 'glm47', 'qwen25', 'hermes'). Used by the SGLang native adapter to parse tool calls from raw model output. |
JSON Schema
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "#/components/schemas/LLMConfig",
"title": "LLMConfig",
"properties": {
"model": {
"type": "string",
"title": "Model",
"description": "LLM model name. "
},
"display_name": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Display Name",
"description": "A human-friendly display name for the model."
},
"model_endpoint_type": {
"type": "string",
"enum": [
"openai",
"anthropic",
"google_ai",
"google_vertex",
"azure",
"groq",
"ollama",
"webui",
"webui-legacy",
"lmstudio",
"lmstudio-legacy",
"lmstudio-chatcompletions",
"llamacpp",
"koboldcpp",
"vllm",
"hugging-face",
"minimax",
"mistral",
"together",
"bedrock",
"deepseek",
"xai",
"zai",
"zai_coding",
"baseten",
"fireworks",
"openrouter",
"chatgpt_oauth"
],
"title": "Model Endpoint Type",
"description": "The endpoint type for the model."
},
"model_endpoint": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Model Endpoint",
"description": "The endpoint for the model."
},
"provider_name": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Provider Name",
"description": "The provider name for the model."
},
"provider_category": {
"anyOf": [
{
"$ref": "#/components/schemas/ProviderCategory"
},
{
"type": "null"
}
],
"description": "The provider category for the model."
},
"model_wrapper": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Model Wrapper",
"description": "The wrapper for the model."
},
"context_window": {
"type": "integer",
"title": "Context Window",
"description": "The context window size for the model."
},
"put_inner_thoughts_in_kwargs": {
"anyOf": [
{
"type": "boolean"
},
{
"type": "null"
}
],
"title": "Put Inner Thoughts In Kwargs",
"description": "Puts 'inner_thoughts' as a kwarg in the function call if this is set to True. This helps with function calling performance and also the generation of inner thoughts.",
"default": false
},
"handle": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Handle",
"description": "The handle for this config, in the format provider/model-name."
},
"temperature": {
"type": "number",
"title": "Temperature",
"description": "The temperature to use when generating text with the model. A higher temperature will result in more random text.",
"default": 1
},
"max_tokens": {
"anyOf": [
{
"type": "integer"
},
{
"type": "null"
}
],
"title": "Max Tokens",
"description": "The maximum number of tokens to generate. If not set, the model will use its default value."
},
"enable_reasoner": {
"type": "boolean",
"title": "Enable Reasoner",
"description": "Whether or not the model should use extended thinking if it is a 'reasoning' style model",
"default": true
},
"reasoning_effort": {
"anyOf": [
{
"type": "string",
"enum": [
"none",
"minimal",
"low",
"medium",
"high",
"xhigh"
]
},
{
"type": "null"
}
],
"title": "Reasoning Effort",
"description": "The reasoning effort to use when generating text reasoning models"
},
"max_reasoning_tokens": {
"type": "integer",
"title": "Max Reasoning Tokens",
"description": "Configurable thinking budget for extended thinking. Used for enable_reasoner and also for Google Vertex models like Gemini 2.5 Flash. Minimum value is 1024 when used with enable_reasoner.",
"default": 0
},
"effort": {
"anyOf": [
{
"type": "string",
"enum": [
"low",
"medium",
"high",
"max"
]
},
{
"type": "null"
}
],
"title": "Effort",
"description": "The effort level for Anthropic models that support it (Opus 4.5, Opus 4.6). Controls token spending and thinking behavior. Not setting this gives similar performance to 'high'."
},
"frequency_penalty": {
"anyOf": [
{
"type": "number"
},
{
"type": "null"
}
],
"title": "Frequency Penalty",
"description": "Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim. From OpenAI: Number between -2.0 and 2.0."
},
"compatibility_type": {
"anyOf": [
{
"type": "string",
"enum": [
"gguf",
"mlx"
]
},
{
"type": "null"
}
],
"title": "Compatibility Type",
"description": "The framework compatibility type for the model."
},
"verbosity": {
"anyOf": [
{
"type": "string",
"enum": [
"low",
"medium",
"high"
]
},
{
"type": "null"
}
],
"title": "Verbosity",
"description": "Soft control for how verbose model output should be, used for GPT-5 models."
},
"tier": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Tier",
"description": "The cost tier for the model (cloud only)."
},
"parallel_tool_calls": {
"anyOf": [
{
"type": "boolean"
},
{
"type": "null"
}
],
"title": "Parallel Tool Calls",
"description": "Deprecated: Use model_settings to configure parallel tool calls instead. If set to True, enables parallel tool calling. Defaults to False.",
"default": false,
"deprecated": true
},
"response_format": {
"anyOf": [
{
"oneOf": [
{
"$ref": "#/components/schemas/TextResponseFormat"
},
{
"$ref": "#/components/schemas/JsonSchemaResponseFormat"
},
{
"$ref": "#/components/schemas/JsonObjectResponseFormat"
}
],
"discriminator": {
"propertyName": "type",
"mapping": {
"json_object": "#/components/schemas/JsonObjectResponseFormat",
"json_schema": "#/components/schemas/JsonSchemaResponseFormat",
"text": "#/components/schemas/TextResponseFormat"
}
}
},
{
"type": "null"
}
],
"title": "Response Format",
"description": "The response format for the model's output. Supports text, json_object, and json_schema (structured outputs). Can be set via model_settings."
},
"strict": {
"type": "boolean",
"title": "Strict",
"description": "Enable strict mode for tool calling. When true, tool schemas include strict: true and additionalProperties: false, guaranteeing tool outputs match JSON schemas.",
"default": false
},
"return_logprobs": {
"type": "boolean",
"title": "Return Logprobs",
"description": "Whether to return log probabilities of the output tokens. Useful for RL training.",
"default": false
},
"top_logprobs": {
"anyOf": [
{
"type": "integer"
},
{
"type": "null"
}
],
"title": "Top Logprobs",
"description": "Number of most likely tokens to return at each position (0-20). Requires return_logprobs=True."
},
"return_token_ids": {
"type": "boolean",
"title": "Return Token Ids",
"description": "Whether to return token IDs for all LLM generations via SGLang native endpoint. Required for multi-turn RL training with loss masking. Only works with SGLang provider.",
"default": false
},
"tool_call_parser": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"title": "Tool Call Parser",
"description": "SGLang tool call parser name (e.g. 'glm47', 'qwen25', 'hermes'). Used by the SGLang native adapter to parse tool calls from raw model output."
}
},
"type": "object",
"required": [
"model",
"model_endpoint_type",
"context_window"
],
"description": "Configuration for Language Model (LLM) connection and generation parameters.\n\n.. deprecated::\n LLMConfig is deprecated and should not be used as an input or return type in API calls.\n Use the schemas in letta.schemas.model (ModelSettings, OpenAIModelSettings, etc.) instead.\n For conversion, use the _to_model() method or Model._from_llm_config() method."
}