MiMo-V2.6-Pro-RL-mxfp4-q8

mlx-community/MiMo-V2.6-Pro-RL-mxfp4-q8#

This model mlx-community/MiMo-V2.6-Pro-RL-mxfp4-q8 was converted to MLX format from XiaomiMiMo/MiMo-V2.6-Pro-RL using mlx-lm version 0.32.0 (PR #1219).

Quantization#

• MoE expert weights are the original checkpoint's native MXFP4 (4-bit, group size 32), loaded directly without requantization.
• Attention, dense MLP, embeddings and lm_head are 8-bit affine, group size 64.
• 4.339 bits per weight overall, 516 GB on disk.

This is a text-only conversion: the vision and audio encoders and the MTP/DFlash draft weights are not included.

Requirements#

MiMo-V2 support is in mlx-lm PR #1219. Until it is merged, install mlx-lm from that branch:

pip install git+https://github.com/kernelpool/mlx-lm.git@add-mimo-v2

At 1.02T total parameters (42B active) this model does not fit on a single Mac. It is meant to run tensor-parallel across two 512 GB machines with mlx-lm's distributed support, for example:

mlx.launch --backend jaccl --hostfile hosts.json -- \
  python -m mlx_lm.examples.sharded_generate \
  --model mlx-community/MiMo-V2.6-Pro-RL-mxfp4-q8 --prompt "hello" -m 256

See the mlx-lm distributed inference documentation for the hostfile format and backend setup.

Use with mlx#

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/MiMo-V2.6-Pro-RL-mxfp4-q8")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, return_dict=False,
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)

Thinking is enabled by default in the chat template; pass enable_thinking=False to apply_chat_template to disable it. Tool calls use the Qwen3-Coder format (<tool_call><function=...>), which the qwen3_coder tool parser in mlx-lm handles.