Attribute	Value
Base Model	Qwen/Qwen2.5-72B-Instruct
Parameters	72.7B
Architecture	Dense Transformer (GQA, 64 attn heads, 8 KV heads)
Quantization	NVFP4 (W4A4 with FP4 weights and dynamic FP4 activations)
Format	`compressed-tensors` (safetensors)
Precision	FP4 weights (group_size=16), FP8 scales, lm_head unquantized
Approx. Size	~42 GB (down from ~145 GB in BF16)
Context Length	32,768 tokens
License	Qwen License

How to Use

vLLM (recommended)

python

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "enfuse/Qwen2.5-72B-Instruct-NVFP4"
tokenizer = AutoTokenizer.from_pretrained(model_id)

llm = LLM(model=model_id, tensor_parallel_size=2)
sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=512)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain quantum computing in simple terms."},
]

prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
outputs = llm.generate([prompt], sampling_params)
print(outputs[0].outputs[0].text)

Hardware Requirements

Full NVFP4 (W4A4): Requires NVIDIA Blackwell GPU (B200, GB200, RTX 5090) for native FP4 tensor core support
Weight-only FP4: Older GPUs (H100, A100) can load the model but will only apply weight quantization, not activation quantization
Recommended: 2x B200 with tensor parallelism for optimal throughput

Quantization Details

This model was quantized using LLM Compressor (v0.10.0) with the NVFP4 scheme:

Method: Post-training quantization (PTQ) with calibration
Calibration data: 512 samples from HuggingFaceH4/ultrachat_200k (train_sft split)
Sequence length: 2048 tokens
Scheme: NVFP4 -- FP4 weights with per-group (group_size=16) local scales in FP8, dynamic FP4 activations at inference
Excluded layers: lm_head (kept in original precision)

python

from llmcompressor.modifiers.quantization import QuantizationModifier

recipe = QuantizationModifier(
    targets="Linear",
    scheme="NVFP4",
    ignore=["lm_head"],
)

Infrastructure

Quantization was performed on an NVIDIA DGX B200 system:

8x NVIDIA B200 GPUs (183 GB HBM3e each)
2x Intel Xeon Platinum 8570 (224 threads)
2 TiB system RAM
Ubuntu 24.04 LTS, CUDA 13.0, Driver 580.126.09

Evaluation

OpenLLM v1 Benchmarks

Evaluated using lm-evaluation-harness with vLLM backend, following the RedHatAI NVFP4 evaluation methodology.

Evaluated with --apply_chat_template --fewshot_as_multiturn flags, tensor_parallel_size=2 on NVIDIA B200 GPUs.

Benchmark	Metric	n-shot	NVFP4	BF16 Reference
ARC-Challenge	acc_norm	25	70.31	—
GSM8K	exact_match	5	79.53	95.8¹
HellaSwag	acc_norm	10	79.75	—
MMLU	acc	5	83.68	—
TruthfulQA MC2	acc	0	68.68	—
Winogrande	acc	5	71.19	—

¹ Qwen blog reference scores use different eval settings; direct comparison requires identical configurations.

Qwen2.5-72B-Instruct BF16 Reference Scores

Official scores from the Qwen2.5 blog (different eval methodology):

Benchmark	BF16 Score
MMLU-Pro	71.1
GSM8K	95.8
MATH	83.1
HumanEval	86.6
IFEval (strict-prompt)	84.1
GPQA	49.0

About Enfuse

Enfuse builds sovereign AI infrastructure for regulated enterprises. The Enfuse platform provides on-prem LLM orchestration and an App Factory for shipping governed, compliant AI applications on your own infrastructure.

This quantization is part of our ongoing work to make large language models more accessible and efficient for on-premise deployment, where memory efficiency directly impacts what models organizations can run within their own data centers.

Acknowledgments

Qwen Team for the base Qwen2.5-72B-Instruct model
vLLM Project for LLM Compressor
NVIDIA for the NVFP4 quantization format and Blackwell hardware