---
title: MiniMax-M3.1-Flash-Text-4bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/MiniMax-M3.1-Flash-Text-4bit"
md_url: "https://www.modelscope.cn/models/mlx-community/MiniMax-M3.1-Flash-Text-4bit.md"
repository: mlx-community/MiniMax-M3.1-Flash-Text-4bit
last_updated: 2026-10-23
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - minimax_m3_vl
architectures:
  - MiniMaxM3SparseForConditionalGeneration
base_model:
  - MiniMaxAI/MiniMax-M3.1-Flash
base_model_relation: quantized
parameters: 82.9B
tensor_type:
  - U32
  - BF16
  - F32
  - U8
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 10
stars: 2
tags:
  - moe
  - agent
  - coding
  - mlx
---

# MiniMax-M3.1-Flash-Text-4bit

> MiniMax-M3.1-Flash-Text-4bit - mlx-community 在 ModelScope 开源的模型。mlx-community/MiniMax-M3.1-Flash-Text-4bit

mlx-community/MiniMax-M3.1-Flash-Text-4bit 是 ModelScope 魔搭社区上的 82.9B 参数text-generation模型，采用 other 许可，基于 MiniMaxAI/MiniMax-M3.1-Flash 构建。

- **Repository**: mlx-community/MiniMax-M3.1-Flash-Text-4bit
- **License**: other
- **Tasks**: text-generation
- **Parameters**: 82.9B
- **Base model**: MiniMaxAI/MiniMax-M3.1-Flash
- **Tags**: moe, agent, coding, mlx
- **Downloads**: 10
- **Stars**: 2
- **Last updated**: 2026-10-23

Source: https://www.modelscope.cn/models/mlx-community/MiniMax-M3.1-Flash-Text-4bit

---

# mlx-community/MiniMax-M3.1-Flash-Text-4bit

This model [mlx-community/MiniMax-M3.1-Flash-Text-4bit](https://huggingface.co/mlx-community/MiniMax-M3.1-Flash-Text-4bit) was converted to MLX format from [MiniMaxAI/MiniMax-M3.1-Flash](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) using mlx-lm with [ml-explore/mlx-lm#2003](https://github.com/ml-explore/mlx-lm/pull/2003). Refer to the [original model card](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) for more details on the model.

This is a text-only conversion. The vision tower is not included.

The weights keep the original quantization. The routed experts are NVFP4 with a per-expert global scale, the attention, shared expert and dense MLP layers are MXFP8, and the router stays in float32. Greedy output matched the original checkpoint token for token in testing. Peak memory was about 249 GB in a short generation test.

Tested with mlx-lm 0.32.2.

## Use with mlx

```bash
pip install -U mlx-lm
```

```python
from mlx_lm import load, generate

model, tokenizer = load("mlx-community/MiniMax-M3.1-Flash-Text-4bit")

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)
```