This model was converted to MLX format from BAAI/AREX-Base using mlx-vlm version 0.4.4. Refer to the original model card for more details on the model.

Use with mlx

pip install -U mlx-vlm

python -m mlx_vlm.generate --model m-i/AREX-Base-mxfp4_plus --max-tokens 100 --temperature 0.0 --prompt "Describe this image." --image <path_to_image>

Quant Predicate

'[INFO] Quantized model with 4.883 bits per weight.'

pyhton

def qwen35_122b_experts_only_predicate(path: str, module):
    """
    Qwen3.5-122B MoE:
      - Routed experts -> MXFP4
      - Everything else -> keep BF16
    """

    p = path.lower()

    # ------------------------------------------------------------------
    # Routed experts only
    # ------------------------------------------------------------------
    if "switch_mlp" in p:
        # Keep metadata tensors untouched
        if any(x in p for x in (
            "bias",
            "scales",
        )):
            return False

        return {
            "mode": "mxfp4",
            "groupe":32,
        }

    # ------------------------------------------------------------------
    # Keep everything else BF16
    # ------------------------------------------------------------------
    return False

mlx-community/AREX-Base-mixed-mxfp4-bf16

m-i/AREX-Base-mxfp4_plus

Use with mlx

Quant Predicate

mlx-community/
AREX-Base-mixed-mxfp4-bf16