Update model card according to templates/model_card.md using INC-Quantization-Recipes

README.md

@@ -1,98 +1,71 @@
 ---
-pipeline_tag: image-text-to-text
+base_model:
+- MiniMaxAI/MiniMax-M3
 license: other
-license_name: minimax-community
-license_link: LICENSE
+pipeline_tag: text-generation
 library_name: transformers
 tags:
-  - multimodal
-  - moe
-  - agent
-  - coding
-  - video
+- auto-round
+- mxfp8
+- compressed-tensors
 ---
 
-This model is a MXFP4 mixed model of [MiniMax-M3](https://huggingface.co/MiniMaxAI/MiniMax-M3) tuned by [intel/auto-round](https://github.com/intel/auto-round) with llm_compressor format. Please follow the license of the original model.
-- MOE (block_sparse_moe.experts): MXFP4
-- MOE (block_sparse_moe.shared_experts): MXFP8
+## Model Details
 
+This model is an MXFP8 model of [MiniMaxAI/MiniMax-M3](https://huggingface.co/MiniMaxAI/MiniMax-M3) generated by [intel/auto-round](https://github.com/intel/auto-round) with 200 tuning iterations. Please follow the license of the original model.
 
-| Configuration | GPQA Diamond | AIME 2026 | SWE-Bench Mini |
-|---------------|--------------|-----------|----------------|
-| BF16 | 0.9242 | 0.9000 | 0.70 |
-| MXFP8 | 0.9242 | 0.9333 | 0.68 |
-| MXFP4 (Experts) + MXFP8 (Shared Experts) | **0.9444** | **0.9667** | **0.74** |
+- **Quantized (MXFP8)**: linear layers
+- **Kept in BF16**: lm_head / embed_tokens
 
+## How to Use
 
+### Environment
 
-
-- The MXFP8 baseline model is from [MiniMaxAI/MiniMax-M3-MXFP8)](https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8)
-
-- SWE-Bench Mini is evaluated using [MariusHobbhahn/swe-bench-verified-mini](https://huggingface.co/datasets/MariusHobbhahn/swe-bench-verified-mini). SWE-Bench Verified Mini is a compact subset of SWE-Bench Verified containing 50 carefully selected instances. It preserves the performance distribution, pass rate, and difficulty characteristics of the full SWE-Bench Verified benchmark while requiring only approximately **5 GB** of storage, compared to roughly **130 GB** for the full dataset.
-
-- GPQA Diamond and AIME 2026 evaluations are conducted using [EvalScope](https://github.com/modelscope/evalscope)
-
-
-## How to Run Locally
-
-
+```bash
+pip install vllm==0.29.0
 ```
-vllm serve \
-  INCModelSharing/MiniMax-M3-MXFP4-Mixed-CT-AutoRound-Tuning \
-  --tensor-parallel-size 2 \
-  --max-model-len 131072 \
-  --tool-call-parser minimax_m3 \
-  --reasoning-parser minimax_m3 \
-  --enable-auto-tool-choice \
-  --port 8000
-````
 
-```
-curl -s http://127.0.0.1:8000/v1/chat/completions   -H "Content-Type: application/json"   -d '{
-    "model": "INCModelSharing/MiniMax-M3-MXFP4-Mixed-CT-AutoRound-Tuning",
-    "messages": [
-      {"role":"user","content":"2+3=?"}
-    ],
-    "max_tokens": 10,
-    "extra_body": {
-      "chat_template_kwargs": {
-        "enable_thinking": true
-      }
-    }
-  }' | python3 -m json.tool
+### vLLM Usage
+
+```bash
+vllm serve intel-ai/MiniMax-M3-MXFP4-CT-AutoRound \
+    --trust-remote-code
 ```
 
+## Evaluation
 
+| Benchmark | BF16 | MXFP8 | Recovery |
+|---|---:|---:|---:|
+| GSM8K | 94.60 | 94.10 | 99.47% |
+| MMLU | 86.20 | 85.80 | 99.54% |
+| PIQA | 83.10 | 82.90 | 99.76% |
+| HellaSwag | 75.40 | 74.80 | 99.20% |
+| **Average** | **84.83** | **84.40** | **99.49%** |
 
 ## Generate the Model
 
-
-~~~bash
-auto-round ../MiniMax-M3/ \
-  --iters 200 \
-  --scheme MXFP8 \
-  --ignore_layers "vision_tower,lm_head,block_sparse_moe.gate,\
-embed_tokens,self_attn,patch_merge_mlp,multi_modal_projector,\
-mlp.gate_up_proj,mlp.down_proj" \
-  --layer_config '{mlp.experts:{bits:4,data_type:mx_fp}}' \
-  --format llm_compressor \
-  --output_dir ./MiniMax-M3-MXFP4MoE-MXFP8-tuning200 \
-  --device_map auto
-~~~
+```bash
+auto-round \
+    --model_name MiniMaxAI/MiniMax-M3 \
+    --iters 200 \
+    --scheme MXFP8 \
+    --ignore_layers "lm_head,embed_tokens" \
+    --format llm_compressor \
+    --output_dir MiniMax-M3-MXFP4-CT-AutoRound
+```
 
 ## Ethical Considerations and Limitations
 
 The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
-
 Therefore, before deploying any applications of the model, developers should perform safety testing.
 
 ## Caveats and Recommendations
 
 Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.
-
 Here are a couple of useful links to learn more about Intel's AI software:
 
 - [Intel Neural Compressor](https://github.com/intel/neural-compressor)
+- [AutoRound](https://github.com/intel/auto-round)
 
 ## Disclaimer
 
@@ -100,6 +73,15 @@ The license on this model does not constitute legal advice. We are not responsib
 
 ## Cite
 
-@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }
+```
+@article{cheng2023optimize,
+  title={Optimize weight rounding via signed gradient descent for the quantization of llms},
+  author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi},
+  journal={arXiv preprint arXiv:2309.05516},
+  year={2023}
+}
+```
 
 [arxiv](https://arxiv.org/abs/2309.05516) [github](https://github.com/intel/auto-round)
+
+Signed-off-by: Lv, Kaokao <kaokao.lv@intel.com>