Published: 17 September 2026
NVIDIA published DeepSeek-V4.1-Flash-NVFP4 on 16 September 2026, a four-bit rebuild of DeepSeek’s DeepSeek-V4.1-Flash that runs on four GB300 GPUs. Across the six benchmarks NVIDIA reports, accuracy shifts by no more than 1.5 points in either direction against the original. This is a repackaging of someone else’s model rather than a new one, and NVIDIA says so plainly on the model card.
What is DeepSeek-V4.1-Flash-NVFP4?
It is DeepSeek-V4.1-Flash with its experts converted to a different four-bit number format. Quantization shrinks the numbers inside a model so it needs less memory and less time per answer. DeepSeek already shipped four-bit weights in a format called MXFP4. NVIDIA converted them to NVFP4, its own four-bit format, which Blackwell GPUs handle directly in hardware.
Only part of the model changed. NVIDIA converted the routed experts, the specialist sub-networks a mixture-of-experts model switches between for each word. That covers 384 experts across 40 layers. Attention, the shared experts, the vision components and the Engram memory tables keep the precision DeepSeek shipped.
One detail cuts against expectations: the file got bigger, not smaller. NVFP4 stores its scaling factors in finer groups, which pushes the checkpoint from roughly 476 GiB to 492 GiB. You are trading a little disk for hardware-native maths.
DeepSeek-V4.1-Flash-NVFP4 benchmarks against the MXFP4 original
NVIDIA ran six benchmarks against DeepSeek’s own MXFP4 checkpoint. The takeaway is that the conversion costs you nothing you can measure.
| What is measured | NVFP4 (this build) | MXFP4 (DeepSeek original) |
|---|---|---|
| GPQA Diamond (hard science questions) | 91.288 | 91.035 |
| AA-LCR (long-context recall) | 78.438 | 78.563 |
| SciCode (scientific coding) | 55.843 | 54.401 |
| IFBench (following instructions) | 77.267 | 76.667 |
| MMMU-Pro (reading images and charts) | 73.699 | 74.046 |
| Terminal-Bench 2.1 (terminal agent tasks) | 82.16 | 81.60 |
Both columns are NVIDIA’s own measurements, taken with vLLM on GB300 hardware. GPQA Diamond and AA-LCR used 16 repeats, IFBench used five, and Terminal-Bench used a separate agent setup with eight attempts per task. The gaps are small enough to sit inside normal run-to-run variation. That is the claim NVIDIA is making, and it is a modest one: nothing broke.
What you need to run DeepSeek-V4.1-Flash-NVFP4
- Licence: MIT, cleared by NVIDIA for commercial and non-commercial use
- Download: Hugging Face, under nvidia/DeepSeek-V4.1-Flash-NVFP4
- Hardware: four NVIDIA GB300 GPUs, Blackwell architecture only
- Runtimes: vLLM and SGLang, both tested by NVIDIA on GB300
- Size: about 492 GiB, up from roughly 476 GiB for the source checkpoint
- Context: up to 1 million tokens, text and images in, text out
The model itself is large but sparse: 552B parameters in the backbone, of which 8B are active while it reads your prompt and 16B while it writes, plus a separate 196B memory store. That is why four GPUs are enough for a model of this size.
What NVIDIA is not saying about DeepSeek-V4.1-Flash-NVFP4
The card publishes accuracy and nothing else. There is no throughput figure, no tokens per second, no time to first token and no memory saving. Speed is the entire reason to move from MXFP4 to NVFP4 on Blackwell, and that number is absent. You are being told the conversion is safe, not that it is fast.
Two smaller gaps matter if you are planning a rollout. NVIDIA kept the DSpark tensors used for speculative decoding, a trick that drafts several tokens at once to speed up output, but states it did not test that path. And the vLLM example on the card is text only, so serving images needs a separate multimodal configuration, even though image input is half of what DeepSeek-V4.1-Flash is for.
What this means
If you already serve DeepSeek-V4.1-Flash on GB300 hardware, download this and time it yourself this week. The accuracy question is settled, so the only thing left to measure is the speed gain, and NVIDIA has left that for you to find. For a platform team running one large open-weights model across four GPUs, a free swap with no quality cost and a plausible throughput win is worth an afternoon of benchmarking.
For everyone else this is safe to ignore. It needs GB300 class Blackwell hardware, which rules out most teams running open weights on their own servers, and it adds no capability whatsoever. Treating it as a DeepSeek release would be a mistake. The interesting part is the timing. DeepSeek shipped V4.1-Flash on 10 September and NVIDIA’s four-bit build landed six days later; GLM-5.3-Flash took seven days in August. Frontier open weights now arrive Blackwell-ready inside a week, which shortens the gap between a Chinese lab publishing a model and a European enterprise being able to serve it on hardware it already owns. That trend matters more than this particular checkpoint does.
For more information, visit the official announcement of DeepSeek-V4.1-Flash-NVFP4 in the NVIDIA model card.
