DeepSeek-R1-Distill-Qwen-32B-NVFP4

Model Overview

• Model Architecture: DeepSeek-R1-Distill-Qwen-32B
  • Input: Text / Image
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP4
  • Activation quantization: FP4
• Release Date: 7/30/25
• Version: 1.0
• Model Developers: RedHatAI

This model is a quantized version of DeepSeek-R1-Distill-Qwen-32B. It was evaluated on a several tasks to assess the its quality in comparison to the unquatized model.

Model Optimizations

This model was obtained by quantizing the weights and activations of DeepSeek-R1-Distill-Qwen-32B to FP4 data type, ready for inference with vLLM>=0.9.1 This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 25%.

Only the weights of the linear operators within transformers blocks are quantized using LLM Compressor.

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

Model Usage Code

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/DeepSeek-R1-Distill-Qwen-32B-NVFP4"
number_gpus = 2

sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

This model was created by applying LLM Compressor with calibration samples from neuralmagic/calibration dataset, as presented in the code snipet below.

Model Creation Code

Evaluation

This model was evaluated on the well-known OpenLLM v1 and HumanEval_64 benchmarks using lm-evaluation-harness. The Reasoning evals were done using ligheval.

Accuracy

Category	Metric	DeepSeek-R1-Distill-Qwen-32B	DeepSeek-R1-Distill-Qwen-32B NVFP4	Recovery
OpenLLM V1	arc_challenge	63.48	62.12	97.86
	gsm8k	86.88	88.32	101.66
	hellaswag	83.51	82.38	98.65
	mmlu	80.97	80.42	99.32
	truthfulqa_mc2	56.82	55.75	98.12
	winogrande	75.93	75.14	98.96
	Average	74.60	74.02	99.23
Reasoning	AIME24 (0-shot)	72.41	62.07	85.69
	AIME25 (0-shot)	58.62	62.07	105.89
	GPQA (Diamond, 0-shot)	68.02	65.48	96.27
	Average	66.35	63.21	95.95
Coding	HumanEval_64 pass@2	90.00	89.32	99.24

Reproduction

The results were obtained using the following commands:

Model Evaluation Commands

OpenLLM v1

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/DeepSeek-R1-Distill-Qwen-32B-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks openllm \
  --batch_size auto

HumanEval_64

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/DeepSeek-R1-Distill-Qwen-32B-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks humaneval_64_instruct \
  --batch_size auto

LightEval

# --- model_args.yaml ---
cat > model_args.yaml <<'YAML'
model_parameters:
  model_name: "RedHatAI/DeepSeek-R1-Distill-Qwen-32B-NVFP4"
  dtype: auto
  gpu_memory_utilization: 0.9
  tensor_parallel_size: 2
  max_model_length: 40960
  generation_parameters:
    seed: 42
    temperature: 0.6
    top_k: 50
    top_p: 0.95
    min_p: 0.0
    max_new_tokens: 32768
YAML

lighteval vllm model_args.yaml \
  "lighteval|aime24|0,lighteval|aime25|0,lighteval|gpqa:diamond|0" \
  --max-samples -1 \
  --output-dir out_dir