Nvidia published an NVFP4-quantized build of DeepSeek-V4-Pro-0813 on Hugging Face on Aug. 27, converting DeepSeek’s 1.65 trillion parameter Mixture-of-Experts model to a lower-precision format tuned for its own Blackwell B200 GPUs. DeepSeek trained and released the underlying model. Nvidia did not build it; Nvidia repackaged an already-open checkpoint so it runs cheaper on Nvidia hardware.
That distinction matters more than it looks. A chip vendor publishing quantized versions of a rival lab’s open weights is not a research contribution. It is a distribution play. The cheaper a model becomes to serve on a given chip, the more inference traffic that chip captures, regardless of who trained the weights in the first place.
DeepSeek-V4-Pro-0813 activates 49 billion of its 1.65 trillion parameters per token. It pairs a hybrid attention design, combining a sparse variant with a heavily compressed one, with a separate technique Nvidia’s card labels Manifold-Constrained Hyper-Connections. The card names reasoning, agentic tool use, and software engineering as the intended use cases, and the checkpoint carries DeepSeek’s DSpark speculative-decoding module through unchanged. Nvidia’s card adds a caution here: this release’s validation did not test speculative decoding on the new checkpoint, so any team turning it on is trying a combination Nvidia has not confirmed works.
Nvidia used its own Model Optimizer software, version 0.46.0, to quantize only the linear operators inside the model’s Mixture-of-Experts transformer blocks (the rest of the network stays at its original precision). Calibration ran on the CNN/DailyMail news dataset and Nvidia’s own Nemotron-Post-Training-Dataset-v2. The card ships under an MIT license for commercial and non-commercial use, targets the SGLang inference engine, and lists Blackwell B200 as the only hardware Nvidia tested it on, with Linux as the preferred operating system.
Nvidia’s own benchmark table compares the NVFP4 build against an unquantized FP8 baseline across six evaluations, including GPQA Diamond, AA-LCR, and Terminal-Bench Hard. The two versions land within roughly a point of each other on every measure, sometimes higher, sometimes lower. That comparison is Nvidia grading its own quantization of a model it did not train, run on the inference stack Nvidia prefers. DeepSeek has not published its own confirmation of those NVFP4 numbers, and no independent benchmark accompanies the release.
The training data behind DeepSeek-V4-Pro-0813 itself is listed as undisclosed on the card, a gap Nvidia cannot close because Nvidia never trained the model. What Nvidia does control is the inference layer: pairing a fast-moving open model with the numeric format its newest chips handle best.
Teams already serving DeepSeek’s models should treat this as a benchmark candidate for Blackwell hardware, not a drop-in upgrade, especially around the unverified speculative-decoding path. Everyone else should read it as a signal that the contest over which GPU vendor captures open-model inference volume is now being fought one quantized checkpoint at a time.
Based on the Hugging Face model card for nvidia/DeepSeek-V4-Pro-0813-NVFP4, published by Nvidia on Aug. 27, 2026.
