Qwen3.5-397B-A17B-NVFP4-Plus-AutoRound

Model Details

This model is an NVFP4+ (also called uNVFP4) model of Qwen/Qwen3.5-397B-A17B generated by intel/auto-round using model-free optimized RTN. Please follow the license of the original model.

NVFP4+ uses FP4 E2M1 elements for weights and activations, an FP8 E5M3 block scale with group size 16, and no tensor-wise global scale. Activations use dynamic per-block scaling.

• Quantized (NVFP4+): all Linear layers in model.language_model.layers and mtp.layers, including routed experts, shared experts, attention projections, linear-attention projections, and the MTP FC/attention layers
• Kept in BF16: router (mlp.gate), mlp.shared_expert_gate, in_proj_a / in_proj_b, lm_head, embeddings, and the vision tower
• Kept in F32: linear_attn.norm
• Not quantized: KV cache and attention softmax / SDPA

How to Use

Environment

The checkpoint uses the auto_round:llm_compressor_nvfp4_e5m3 packing format. Use an inference stack whose compressed-tensors / vLLM build supports NVFP4+ with E5M3 block scales; the checkpoint was produced with auto-round 0.16.0.

vLLM Usage

vllm serve intel-ai/Qwen3.5-397B-A17B-NVFP4-Plus-AutoRound

Evaluation

Recovery is the quantized score divided by the BF16 baseline.

Average loss versus BF16: 0.96%.

Generate the Model

auto-round Qwen/Qwen3.5-397B-A17B \
    --model_free \
    --scheme NVFP4_E5M3 \
    --ignore_layers "visual,lm_head,embed_tokens,mlp.gate,mlp.shared_expert_gate,in_proj_a,in_proj_b" \
    --format auto_round \
    --output_dir ./Qwen3.5-397B-A17B-NVFP4-Plus

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
• AutoRound

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 github

Signed-off-by: Kaokao, Lv [email protected]