MLX-Textgen

An OpenAI-compatible API LLM engine with smart prompt caching, batch processing, structured output with guided decoding, and function calling for models using MLX

MLX-Textgen is a light-weight LLM serving engine that utilize MLX and a smart KV cache management system to make your LLM generation more seamless on your Apple silicon machine. It features:

• Multiple KV-cache slots to reduce the needs of prompt processing
• Structured text generation with json schemas, regex, or context free grammar
• Batch inferencing with multiple prompts
• Multiple models serving with Fastapi
• Common OpenAI API endpoints: /v1/models, /v1/completions, /v1/chat/completions

It is built with:

1. mlx-lm
2. mlx-vlm
3. Outlines
4. FastAPI

Updates

• 2025-06-21: Some vision models supported with mlx-vlm integration. Tested with Gemma 3 family models and Mistral Small 3.1.
• 2025-06-21: Reasoning parser supported for reasoning models with deepseek_r1 parser.
• 2025-06-21: Breaking changes due to new vision model support and code restructuring. Run mlx_texgen createconfig to create a new config file.
• 2025-06-21: Quantising models will need to be done manually with mlx-lm or mlx-vlm.

Installing MLX-Textgen

MLX-textgen can be easily installed with pip:

pip install mlx-textgen

Usage

1. Serving a single model

You can quickly set up a OpenAI API server with a single command.

mlx_textgen serve --model-path mlx-community/gemma-3-4b-it-8bit --port 5001

2. Serving a multiple models server

Create a config file template and add as many model as you like.

mlx_textgen createconfig --num-models 2

It will generate a file called model_config.yaml. Edit this file for the models you want to serve.

model_configs:
- model_id_or_path: /path/to/model_0
  tokenizer_repo_or_path: null
  model_kwargs: null
  tokenizer_kwargs: null
  model_name: null
  enable_cache: true
  preprocess_batch_size: 512
  extra_stop_words: null
  reasoning_parser: null
  default_template: null
- model_id_or_path: /path/to/model_1
  tokenizer_repo_or_path: null
  model_kwargs: null
  tokenizer_kwargs: null
  model_name: null
  enable_cache: true
  preprocess_batch_size: 512
  extra_stop_words: null
  reasoning_parser: null
  default_template: null
host: 127.0.0.1
port: 5001
api_keys: null
min_tokens: 20
max_reprocess_tokens: 250
replace_threshold: 0.95
max_capacity: 50
use_reasoning_content: false

Then start the engine:

mlx_textgen serve --config-file ./model_config.yaml

3. More engine arguments

You can check the details of other engine arguments by running:

mlx_textgen serve --help

You can specify the number of cache slots for each model, minimum number of tokens to create a cache file, and API keys etc.

Features

1. Multiple KV cache slots support

All the KV cache are stored on disk. Therefore, unlike other LLM serving engine, a newly created KV cache will not overwrite the existing KV cache. This works better for agentic workflows where different types of prompts are being used frequently without losing previous cache for a long prompt.

2. Guided decoding with Regex, Json schema, and Grammar

You can pass your guided decoding argument guided_json, guided_choice, guided_regex, or guided_grammar as extra arguments and create structured text generation in a similar fashion to vllm.

3. Batch inference support

Batch inference is supported for multiple prompts or multiple generations for a single prompt. Just pass a list of prompts to the prompt argument to the /v1/completions endpoint or n=2 (or more than 2) to the /v1/chat/completions or v1/completions endpoints for batch inferencing.

4. Function calling support

Function calling with the /v1/chat/completions is supported. Simply use the tools and tool_choice arguments to supply lists of tools. There are three modes of using function calling:

1. tool_choice="auto": The model will decide if tool calling is needed based on the conversation. If a tool is needed, it will pick the appropriate tool and generate the arguments. Otherwise, it will only response with normal text.
2. tool_choice="required": One of the given tools must be selected by the model. The model will pick the appropriate tool and generate the arguments.
3. tool_choice={"type": "function", "function": {"name": "<selected tool name>"}}: The model will generate the arguments of the selected tools.

If function calling is triggered, the call arguments will be contained in the tool_calls attribute in the choices element in the response. The finish_reason will be tool_calls.

from openai import OpenAI

tools = [{
  "type": "function",
  "function": {
    "name": "get_current_weather",
    "description": "Get the current weather in a given location",
    "parameters": {
      "type": "object",
      "properties": {
        "location": {
          "type": "string"
        },
        "unit": {
          "type": "string",
          "default": "celsius"
        }
      },
      "required": ["location"]
    }
  }
}]

client = OpenAI(api_key='Your API Key', base_url='http://localhost:5001/v1/')

output = client.chat.completions.create(
    model='model_name',
    messages=[
        dict(role='user', content='What is the current weather in London?')
    ],
    max_tokens=256,
    tools=tools,
    tool_choice='auto',
    stream=False
).choices[0].model_dump()

# output: 
# {'finish_reason': 'tool_calls',
#  'index': 0,
#  'logprobs': None,
#  'message': {'content': None,
#   'role': 'assistant',
#   'function_call': None,
#   'tool_calls': [{'id': 'call_052c8a6b',
#     'function': {'arguments': '{"location": "London", "unit": "celsius" }',
#      'name': 'get_current_weather'},
#     'type': 'function',
#     'index': 0}]}}

If tool_choice="none" is passed, the list of tools provided will be ignored and the model will only generate normal text.

5. Multiple LLMs serving

Only one model is loaded on ram at a time, but the engine leverage MLX fast module loading time to spin up another model when it is requested. This allows serving multiple models with one endpoint.

License

This project is licensed under the terms of the MIT license.