deepseek nvfp4: NVIDIA Shrinks V4.1 Flash to 400 GiB

The deepseek nvfp4 checkpoint is NVIDIA's new way to squeeze DeepSeek's giant open model onto its latest data-center GPUs. Published on Hugging Face on October 30 and covered by Open Model Weights on November 1, NVIDIA's DeepSeek-V4.1-Flash-1017-NVFP4 shrinks the checkpoint to about 400 GiB from roughly 475 GiB for the original package.

What actually happened

DeepSeek released V4.1 Flash 1017 earlier, and Daily Launch's November 1 list showed it with about 4,100 likes on Hugging Face. We covered it in deepseek v4 1 flash 1017. NVIDIA's team then ran it through its Model Optimizer toolkit to produce a four-bit version.

The model card describes a natively multimodal mixture-of-experts model that accepts text and images, outputs text, and supports up to one million tokens of context. NVIDIA says the release is ready for commercial and non-commercial use under the MIT license.

Open Model Weights stresses the key point: this changes how the weights are stored and served, not what the model is. The download and hardware needs change. The underlying intelligence does not.

What the deepseek nvfp4 format changes

NVFP4 is NVIDIA's four-bit floating-point format. In this checkpoint, the expert projections in the mixture-of-experts layers use NVFP4 for both weights and activations, with a group size of 16, according to Open Model Weights. The Engram lookup tables use a different four-bit format, MXFP4.

Not everything is shrunk. Attention, shared experts and vision components keep their original precision. So this is not a uniform all-FP4 model.

The publisher says converting the ordinary expert weights preserved their values exactly, while the Engram conversion is lossy. Activation scales also needed calibration and a few fallback values. Those details matter more than any marketing word like "lossless."

Smaller files are great. But 400 GiB is still a data-center download, not something for a gaming laptop.

Who can actually run it

The model card says stock vLLM images cannot load the MXFP4 Engram tables. You need a vLLM build that includes a specific pull request. NVIDIA's own benchmarks used vLLM v0.31.0 with those changes on four GB300 GPUs. SGLang has not been validated.

That limits the audience to teams with Blackwell hardware and the know-how to run patched inference stacks. For them, a 75 GiB smaller checkpoint means more room for context and batch size on the same GPUs.

NVIDIA published comparison scores against the original model, but they come from its own setup. Open Model Weights says it has not reproduced them, and neither have we.

Why it matters for the open-model race

The deepseek nvfp4 release also shows how quickly the ecosystem moves around a popular open model. Within days of a DeepSeek update, a hardware vendor shipped an optimized build with its own serving notes. That speed is good for users, but it also means more versions to track, each with slightly different behavior and requirements.

Open-weight models keep getting bigger, and serving cost is now the main barrier. Vendor-made quantized checkpoints lower that barrier for large users and tie popular open models more tightly to specific hardware.

This fits a broader trend of hardware makers packaging open models for their own chips. We saw NVIDIA's own model work in nemotron 3 5 transcribe. Expect more of these releases as labs ship larger mixture-of-experts models.

For crypto-adjacent builders running AI agents, cheaper serving of strong open models can mean lower costs for on-chain analytics and trading bots. It does not change the need to test outputs before trusting them with funds.

What to do as a reader

If you run Blackwell GPUs, read the NVIDIA model card and check the vLLM requirement before downloading 400 GiB.

Test the deepseek nvfp4 build on your own tasks against the original weights. Publisher benchmarks are a starting point, not a guarantee for your workload. If you lack the hardware, hosted APIs remain the practical way to use the model, and they spare you the patched-runtime headaches and the large upfront hardware bill entirely.