Skip to main content
MLX-VLM registers four CLI entry points when you install the package. Each one maps directly to a Python module’s main() function.
CommandModulePurpose
mlx_vlm.generatemlx_vlm.generateOne-shot text generation from text, images, or audio
mlx_vlm.chat_uimlx_vlm.chat_uiLaunch a Gradio chat interface
mlx_vlm.convertmlx_vlm.convertConvert and quantize Hugging Face checkpoints
mlx_vlm.servermlx_vlm.serverStart an OpenAI-compatible HTTP server

​
mlx_vlm.generate

The primary inference command. Supports text-only, image, audio, and multi-modal inputs.

​
Examples

mlx_vlm.generate \
  --model mlx-community/Qwen2-VL-2B-Instruct-4bit \
  --max-tokens 100 \
  --prompt "Hello, how are you?"

​
Thinking budget

For reasoning models such as Qwen3.5, you can cap the number of tokens spent inside the thinking block: 
mlx_vlm.generate \
  --model mlx-community/Qwen3.5-2B-4bit \
  --enable-thinking \
  --thinking-budget 50 \
  --thinking-start-token "<think>" \
  --thinking-end-token "</think>" \
  --prompt "Solve 2+2"
When the budget is exceeded, the model is forced to emit \n</think> and transition to its answer. If --enable-thinking is set but the model’s chat template does not support it, the budget is applied only when the model generates the start token on its own.

​
Activation quantization (CUDA)

Models quantized with mxfp8 or nvfp4 require activation quantization on NVIDIA GPUs. Use the -qa shorthand or the full flag: 
mlx_vlm.generate \
  --model /path/to/mxfp8-model \
  --image /path/to/image.jpg \
  --prompt "Describe this image" \
  -qa
On Apple Silicon (Metal), mxfp8 and nvfp4 models work without the -qa flag.

​
Flag reference

FlagTypeDefaultDescription
--modelstringmlx-community/nanoLLaVA-1.5-8bitHugging Face repo ID or local model path
--adapter-pathstringNonePath to LoRA adapter weights
--imagestring (one or more)NoneURL(s) or local path(s) of images to process
--audiostring (one or more)NoneURL(s) or local path(s) of audio files to process
--resize-shapeint (one or two values)NoneResize images to this shape before processing
--promptstring"What are these?"Text prompt sent to the model
--systemstringNoneSystem message prepended to the conversation
--max-tokensint256Maximum number of tokens to generate
--temperaturefloat0.0Sampling temperature; 0 uses argmax (greedy)
--max-kv-sizeintNoneMaximum KV cache size for long-context prompts
--kv-bitsintNoneQuantize the KV cache to this many bits
--kv-group-sizeint64Group size used when quantizing the KV cache
--quantized-kv-startint5000Token index at which KV cache quantization begins
--prefill-step-sizeint2048Tokens processed per prefill chunk; lower values reduce peak memory
--enable-thinkingflagFalseActivate thinking mode in the chat template
--thinking-budgetintNoneMaximum tokens allowed inside a thinking block
--thinking-start-tokenstring<think>Token that opens a thinking block
--thinking-end-tokenstring</think>Token that closes a thinking block
--quantize-activations / -qaflagFalseEnable activation quantization for mxfp8/nvfp4 models
--processor-kwargsJSON string{}Extra kwargs forwarded to the processor (e.g. '{"cropping": false}')
--eos-tokensstring (one or more)NoneAdditional end-of-sequence tokens
--skip-special-tokensflagFalseOmit special tokens from the decoded output
--chatflagFalseEnter multi-turn chat mode
--verboseflagTruePrint tokens and timing statistics as they are generated
--trust-remote-codeflagFalseAllow execution of remote code when loading the model
--revisionstring"main"Model branch, tag, or commit to use
--force-downloadflagFalseRe-download the model even if it is already cached

​
mlx_vlm.chat_ui

Launches an interactive Gradio chat interface in your browser. 
mlx_vlm.chat_ui --model mlx-community/Qwen2-VL-2B-Instruct-4bit
The UI loads the model at startup and exposes a chat window that accepts text and image inputs. You can switch models from within the interface; the previous model is unloaded from memory before the new one is loaded.
The gradio package is an optional dependency. Install it with pip install 'mlx-vlm[ui]' before using this command.
FlagTypeDefaultDescription
--modelstringqnguyen3/nanoLLaVAHugging Face repo ID or local path of the model to load at startup

​
mlx_vlm.convert

Converts a Hugging Face Vision Language Model checkpoint to MLX format with optional quantization. 
mlx_vlm.convert \
  --hf-path mlx-community/Qwen2-VL-2B-Instruct \
  --mlx-path ./Qwen2-VL-2B-Instruct-4bit \
  -q
See the Model conversion guide for the full flag reference and quantization options.

​
mlx_vlm.server

Starts an OpenAI-compatible FastAPI server. See the REST API server page for the complete reference. 
mlx_vlm.server --port 8080

Build docs developers (and LLMs) love

Get started for freeTalk to us