NVIDIA has released DeepSeek-V4-Pro-0813 as an NVFP4 checkpoint, integrating a DSpark-Draft-Head. This release features MXFP4 experts from the draft module losslessly mapped to NVFP4, ensuring consistent quantization between target and draft. The model is designed to run on Blackwell hardware utilizing vLLM. AI
IMPACT This release integrates advanced quantization techniques and is optimized for NVIDIA's Blackwell hardware, potentially improving performance for AI workloads.
RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
Read on Mastodon — mastodon.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →