Nvidia/DeepSeek-V4-Pro-0813-NVFP4
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
NVIDIA has released DeepSeek-V4-Pro-0813-NVFP4, an NVFP4-quantized version of DeepSeek's 1.65-trillion-parameter Mixture-of-Experts model (49B active), published on Hugging Face under the MIT license. Only the MoE linear layers are quantized, and the checkpoint is built for Blackwell B200 with SGLang or vLLM, supporting a 1-million-token context.