NVIDIA DeepSeek-V4-Pro-0813-NVFP4 is a quantized release of DeepSeek AI’s autoregressive Mixture-of-Experts model, intended for advanced reasoning, agentic tasks, tool use, coding, mathematics, and enterprise assistants. It has 1.65 trillion parameters in total, with 49 billion active, and combines a Transformer architecture with compressed sparse and heavily compressed attention. It accepts text, supports multi-turn conversations and contexts up to one million tokens, and can generate structured JSON, call tools, and expose reasoning content when enabled. The model is released under the MIT license for commercial and non-commercial use; its card notes that NVIDIA did not develop the underlying DeepSeek model.
Model Optimizer converts the weights and activations of linear operators in the Transformer’s MoE blocks to NVFP4; DeepSeek’s DSpark speculative-decoding heads remain unchanged. Calibration uses 512 samples each from CNN/DailyMail and NVIDIA’s Nemotron post-training dataset, with sequences capped at 4,096 tokens to better capture long-context activation ranges. The resulting checkpoint is designed for SGLang or vLLM on NVIDIA Blackwell B200 GPUs, with Linux listed as the preferred operating system. Evaluations cover graduate-level science questions, long-context recall, telecom tool use, scientific coding, instruction following, and terminal-based software engineering, but the provided description does not report benchmark scores. The card also emphasizes that deployment requires additional testing against the intended use case.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.