Noesion Radarnoesion.ai
Release2026-08-31

NVIDIA published a version of DeepSeek's V4-Pro model rebuilt to use its own compressed number format, NVFP4, for faster running on Blackwell B200 graphics chips. Benchmark scores across reasoning, coding and tool-use tests are within about a point of the original in both directions. It runs only on the SGLang serving engine, and DeepSeek's built-in speed-up module was left untested in this release.

What changed

DeepSeek-V4-Pro-0813 shipped its routed experts in the MXFP4 number format.

What it unlocks

Running DeepSeek's 1.65-trillion-parameter model on NVIDIA Blackwell B200 hardware in NVIDIA's own NVFP4 format with near-identical benchmark scores.

  • 1.65T params, 49B activated
  • 1M token context
  • 88.42 vs 88.51 GPQA Diamond
  • 384 experts, 8-way parallel

What you need to act on it

  • NVIDIA Blackwell B200 GPUs
  • SGLang runtime on Linux
  • 8-way model parallelism

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.