NVIDIA publishes NVFP4 DeepSeek-V4.1-Flash-1017 checkpoint at about 400 GiB for Blackwell GPUs

NVIDIA has published an NVFP4 quantized version of DeepSeek AI's DeepSeek-V4.1-Flash-1017 on Hugging Face. The checkpoint is about 400 GiB, compared with about 475 GiB for the source model. NVIDIA lists the Hugging Face release date as 10/30/2026 and the license as MIT.

NVIDIA states the model is not owned or developed by NVIDIA. It was quantized with Model Optimizer (nvidia-modelopt v0.48.0rc0) from a model built by DeepSeek AI.

Base model

DeepSeek-V4.1-Flash-1017 is a natively multimodal Mixture-of-Experts model with a Causal Encoder-Decoder architecture and Compressed Sparse Attention 2. It supports contexts of up to one million tokens and accepts text and image input. Output is text.

• Parameters: 552B backbone in total, 8B activated during prefill and 16B during decode, plus 196B Engram conditional memory.
• Training data: listed as undisclosed.
• Pricing: not yet disclosed.

What NVIDIA changed

The conversion moves the routed MoE experts from source MXFP4 to NVFP4 weights and activations (W4A4) with group size 16. The converted projections are w1, w2 and w3 for 384 experts across 40 layers. The Engram lookup tables go from source FP8 to MXFP4. Attention, shared experts, vision, MTP and other excluded components keep their source precision, including MXFP8 where applicable.

All 16,986,931,200 expert weight blocks passed lossless conversion. The Engram conversion is lossy. NVIDIA notes that lossless weight conversion does not imply identical inference outputs. Activation scales were calibrated, and 24 of 46,080 projection entries used fallback scales.

Calibration used cnn_dailymail and Nemotron-Post-Training-Dataset-v2 (stem, chat, math and code subsets), with 512 samples per dataset (1,024 total), sequence length 512 and batch size 4. The checkpoint ships in 48 safetensors shards.

Benchmarks (percent, run with vLLM)

Precision	GPQA Diamond	AA-LCR	SciCode	IFBench	MMMU-Pro	Terminal-Bench 2.1
MXFP4 (source)	92.77	80.41	57.58	76.33	78.38	81.81
NVFP4 + MXFP4 Engram	92.93	80.19	57.32	76.47	78.73	81.11

The NVFP4 checkpoint scores higher on GPQA Diamond, IFBench and MMMU-Pro, and lower on AA-LCR, SciCode and Terminal-Bench 2.1. Requests used temperature=1.0, top_p=1.0, reasoning_effort=max and max_new_tokens=262144. GPQA used 16 repeats, AA-LCR used 16 repeats in each of two runs, IFBench used 5, and SciCode is the mean of 8 runs. Terminal-Bench 2.1 used a separate agentic configuration with eight trajectories per task, and the score is the mean of two runs.

Deployment requirements

• Hardware: NVIDIA Blackwell. Results were produced on four GB300 GPUs with tensor parallelism of 4. Linux is the preferred OS.
• Runtime: vLLM only. The checkpoint stores Engram tables as MXFP4, which stock vLLM images cannot load. It requires a build that includes vllm-project/vllm#61530. Reported results used vllm/vllm-openai:v0.31.0 with that pull request's Engram changes applied.
• Not validated: SGLang. Speculative decoding (MTP) was not exercised. Text and image input were validated.
• Serving options: the reference command uses the deepseek_v41 tokenizer mode and reasoning parser, FP8 KV cache, max-num-seqs 128 and Engram CPU offload. Tool calling needs --enable-auto-tool-choice and --tool-call-parser deepseek_v41. Thinking at maximum effort is set with chat_template_kwargs {"thinking": true, "reasoning_effort": "max"}.

What this means

This is a packaging release, not a new model. The weights come from DeepSeek AI, and NVIDIA's contribution is a lower-precision checkpoint with reported accuracy close to the MXFP4 source on six benchmarks. Deployment is narrow for now: Blackwell hardware, vLLM, and a pull request that stock images do not include. Teams evaluating it should validate the final package on their intended runtime, as NVIDIA advises, because the lossy Engram conversion and the fallback activation scales mean outputs can differ from the source even where benchmark scores are close.