Qwen3.5-4B-AutoRound-NVFP4-Tuning

Model Details

This model is a NVFP4 (NVIDIA FP4) quantization of Qwen/Qwen3.5-4B generated by TUNING. Please follow the license of the original model.

Quantization Details

Attribute	Value
Base Model	Qwen/Qwen3.5-4B
Quantization Tool	TUNING
Quantization Scheme	NVFP4
Quantized Size	4398 MB

Evaluation Results

Task	Accuracy
hellaswag	0.5292
mmlu	0.7195
mmlu_abstract_algebra	0.5200
mmlu_anatomy	0.7481
mmlu_astronomy	0.8553
mmlu_business_ethics	0.7300
mmlu_clinical_knowledge	0.7774
mmlu_college_biology	0.8472
mmlu_college_chemistry	0.5100
mmlu_college_computer_science	0.6500
mmlu_college_mathematics	0.5600
mmlu_college_medicine	0.7225
mmlu_college_physics	0.5882
mmlu_computer_security	0.8000
mmlu_conceptual_physics	0.8170
mmlu_econometrics	0.6579
mmlu_electrical_engineering	0.7793
mmlu_elementary_mathematics	0.7169
mmlu_formal_logic	0.5476
mmlu_global_facts	0.3600
mmlu_high_school_biology	0.8968
mmlu_high_school_chemistry	0.7537
mmlu_high_school_computer_science	0.8100
mmlu_high_school_european_history	0.8424
mmlu_high_school_geography	0.8737
mmlu_high_school_government_and_politics	0.9326
mmlu_high_school_macroeconomics	0.7641
mmlu_high_school_mathematics	0.4481
mmlu_high_school_microeconomics	0.8992
mmlu_high_school_physics	0.6424
mmlu_high_school_psychology	0.9101
mmlu_high_school_statistics	0.6852
mmlu_high_school_us_history	0.8725
mmlu_high_school_world_history	0.8312
mmlu_human_aging	0.7309
mmlu_human_sexuality	0.8244
mmlu_humanities	0.6349
mmlu_international_law	0.8264
mmlu_jurisprudence	0.8148
mmlu_logical_fallacies	0.7975
mmlu_machine_learning	0.5714
mmlu_management	0.8641
mmlu_marketing	0.9188
mmlu_medical_genetics	0.8400
mmlu_miscellaneous	0.8238
mmlu_moral_disputes	0.7486
mmlu_moral_scenarios	0.4145
mmlu_nutrition	0.7941
mmlu_other	0.7593
mmlu_philosophy	0.7621
mmlu_prehistory	0.7778
mmlu_professional_accounting	0.5993
mmlu_professional_law	0.5372
mmlu_professional_medicine	0.7978
mmlu_professional_psychology	0.7549
mmlu_public_relations	0.7091
mmlu_security_studies	0.7265
mmlu_social_sciences	0.8196
mmlu_sociology	0.8657
mmlu_stem	0.7088
mmlu_us_foreign_policy	0.8600
mmlu_virology	0.5663
mmlu_world_religions	0.8363
piqa	0.7628

How to Use

HF Usage

Step 1: Install AutoRound

pip install auto-round

Step 2: Load and run the quantized model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen3.5-4B-AutoRound-NVFP4-Tuning"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")

# prepare the model input
prompt = "Write a quick sort algorithm."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(**model_inputs, max_new_tokens=512)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]) :].tolist()

content = tokenizer.decode(output_ids, skip_special_tokens=True)
print("content:", content)

VLLM Usage

vllm serve Qwen3.5-4B-AutoRound-NVFP4-Tuning \
    --trust-remote-code \
    --dtype bfloat16 \
    --tensor_parallel_size 1

If you encounter any issues, feel free to open an issue on the AutoRound GitHub repo or provide feedback on the Low-Bit Open LLM Leaderboard.

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:

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

This model is part of the Intel Low-Bit Open LLM Leaderboard initiative.