---
title: AREX-Base-mixed-mxfp4-bf16
canonical_url: "https://www.modelscope.cn/models/mlx-community/AREX-Base-mixed-mxfp4-bf16"
md_url: "https://www.modelscope.cn/models/mlx-community/AREX-Base-mixed-mxfp4-bf16.md"
repository: mlx-community/AREX-Base-mixed-mxfp4-bf16
last_updated: 2026-07-30
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3_5_moe
architectures:
  - Qwen3_5MoeForConditionalGeneration
base_model:
  - BAAI/AREX-Base
base_model_relation: quantized
parameters: 28.3B
tensor_type:
  - BF16
  - U32
  - U8
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - Pytorch
language:
  - en
downloads: 14
stars: 0
tags:
  - agent
  - deep-research
  - reasoning
  - tool-use
  - long-context
  - qwen3.5
  - mixture-of-experts
  - mlx
---

# AREX-Base-mixed-mxfp4-bf16

> AREX-Base-mixed-mxfp4-bf16 - mlx-community 在 ModelScope 开源的模型。m-i/AREX-Base-mxfp4plus

mlx-community/AREX-Base-mixed-mxfp4-bf16 是 ModelScope 魔搭社区上的 28.3B 参数text-generation模型，采用 apache-2.0 许可，基于 BAAI/AREX-Base 构建。

- **Repository**: mlx-community/AREX-Base-mixed-mxfp4-bf16
- **License**: apache-2.0
- **Tasks**: text-generation
- **Parameters**: 28.3B
- **Base model**: BAAI/AREX-Base
- **Tags**: agent, deep-research, reasoning, tool-use, long-context, qwen3.5, mixture-of-experts, mlx
- **Downloads**: 14
- **Stars**: 0
- **Last updated**: 2026-07-30

Source: https://www.modelscope.cn/models/mlx-community/AREX-Base-mixed-mxfp4-bf16

---

# m-i/AREX-Base-mxfp4_plus

This model was converted to MLX format from [`BAAI/AREX-Base`](https://huggingface.co/BAAI/AREX-Base)
using mlx-vlm version **0.4.4**.
Refer to the [original model card](https://huggingface.co/BAAI/AREX-Base) for more details on the model.

## Use with mlx

```bash
pip install -U mlx-vlm
```

```bash
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
```