---
base_model:
- thinkingmachines/Inkling-Small
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- auto-round
- mxfp4
---

## Model Details

This model is an MXFP4 (W4A16, group_size 32) model of [thinkingmachines/Inkling-Small](https://huggingface.co/thinkingmachines/Inkling-Small) generated by [intel/auto-round](https://github.com/intel/auto-round) using RTN. Please follow the license of the original model.

- **Quantized (MXFP4)**: MoE routed experts (`mlp.experts`)
- **Kept in BF16**: vision tower, embeddings, lm_head, norms

## How to Use

### Environment

```bash
pip install vllm==0.29.0
```

### vLLM Usage

```bash
vllm serve intel-ai/Inkling-Small-MXFP4-AutoRound \
    --tokenizer-mode inkling \
    --reasoning-parser inkling \
    --tool-call-parser inkling \
    --enable-auto-tool-choice \
    --trust-remote-code
```

## Evaluation

| Benchmark | Base Model | MXFP4 | Recovery |
|---|---:|---:|---:|
| aime25 | 90.00 | 91.11 | 101.23% |
| gpqa_diamond | 87.37 | 88.55 | 101.35% |
| GSM8K | 97.27 | 96.99 | 99.71% |
| PIQA | 94.47 | 94.52 | 100.05% |
| **Average** | **92.28** | **92.79** | **100.55%** |

*Note: Baseline corresponds to NVFP4 reference evaluated under identical settings.*

## Generate the Model

```bash
auto-round \
    --model_name thinkingmachines/Inkling-Small \
    --scheme BF16 \
    --layer_config '{mlp.experts:{scheme:mxfp4}}' \
    --model_free \
    --format llm_compressor \
    --output_dir Inkling-Small-MXFP4-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

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

## 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}
}
```

[arxiv](https://arxiv.org/abs/2309.05516) [github](https://github.com/intel/auto-round)

Signed-off-by: He, Xin <xin3.he@intel.com>