README.md

---
base_model:
- MiniMaxAI/MiniMax-M2.5
language:
- en
library_name: transformers
license: other
license_name: modified-mit
license_link: https://github.com/MiniMax-AI/MiniMax-M2.5/blob/main/LICENSE
---

# Model Overview

- **Model Architecture:** MiniMaxM2ForCausalLM
  - **Input:** Text
  - **Output:** Text
- **Supported Hardware Microarchitecture:** AMD MI300 MI350/MI355
- **ROCm**: 7.0
- **PyTorch**: 2.8.0
- **Transformers**: 4.57.1
- **Operating System(s):** Linux
- **Inference Engine:** [SGLang](https://docs.sglang.ai/)/[vLLM](https://docs.vllm.ai/en/latest/)
- **Model Optimizer:** [AMD-Quark](https://quark.docs.amd.com/latest/index.html) (v0.11)
  - **Weight quantization:** OCP MXFP4, Static
  - **Activation quantization:** OCP MXFP4, Dynamic


# Model Quantization

The model was quantized from [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) which was converted to bf16 using [QuixiAI/MiniMax-M2.1-bf16/minimax_to_bf16.py](https://huggingface.co/QuixiAI/MiniMax-M2.1-bf16/blob/main/minimax_to_bf16.py) using [AMD-Quark](https://quark.docs.amd.com/latest/index.html). The weights are quantized to MXFP4 and activations are quantized to MXFP4.


**Quantization scripts:**
```
cd Quark/examples/torch/language_modeling/llm_ptq/
export exclude_layers="lm_head *block_sparse_moe.gate* *self_attn*"
python3 quantize_quark.py --model_dir $MODEL_DIR \
                          --quant_scheme mxfp4 \
                          --num_calib_data 128 \
                          --exclude_layers $exclude_layers \
                          --skip_evaluation \
                          --multi_gpu  \
                          --trust_remote_code \
                          --model_export hf_format \
                          --output_dir $output_dir
```
For further details or issues, please refer to the AMD-Quark documentation or contact the respective developers.

# Evaluation
The model was evaluated on gsm8k benchmarks using the [vllm](https://github.com/vllm-project/vllm/tree/v0.13.0) framework.

### Accuracy

<table>
  <tr>
   <td><strong>Benchmark</strong>
   </td>
   <td><strong>MiniMaxAI/MiniMax-M2.5 </strong>
   </td>
   <td><strong>amd/MiniMax-M2.5-MXFP4(this model)</strong>
   </td>
   <td><strong>Recovery</strong>
   </td>
  </tr>
  <tr>
   <td>gsm8k (flexible-extract) 
   </td>
   <td>0.9401
   </td>
   <td>0.9256
   </td>
   <td>98.46%
   </td>
  </tr>
</table>


### Reproduction

The GSM8K results were obtained using the vLLM framework, based on the Docker image `rocm/vllm:rocm7.0.0_vllm_0.11.2_20251210`, and vLLM is installed inside the container with fixes applied for model support.

#### Preparation in container
```
# Reinstall vLLM
pip uninstall vllm -y
git clone https://github.com/vllm-project/vllm.git
cd vllm
git checkout v0.13.0
pip install -r requirements/rocm.txt
python setup.py develop
cd ..
```

#### Launching server
```
VLLM_ROCM_USE_AITER=1 \
VLLM_DISABLE_COMPILE_CACHE=1 \
vllm serve "$MODEL" \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --max-model-len 32768 \
    --port 8899  
```


#### Evaluating model in a new terminal
```
python vllm/tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port 8899 --num-questions 1000 --save-results logs
```


# License
Modifications Copyright(c) 2026 Advanced Micro Devices, Inc. All rights reserved.