Nvidia Blackwell Platform Adapts DeepSeek-V4 Models, Out-of-the-Box Performance Exceeds 150 Tokens/sec/User

Nvidia announced its Blackwell platform now supports DeepSeek-V4-Pro and DeepSeek-V4-Flash models. Developers can deploy via NVIDIA NIM microservices or customize inference using SGLang and vLLM frameworks.

DeepSeek-V4-Pro has 1.6T total parameters and 49B active parameters for advanced reasoning tasks; the Flash version has 284B total and13B active parameters for high-speed scenarios. Both support a 1M token context window and use the MIT open-source license.

Testing shows DeepSeek-V4-Pro delivers over150 tokens/sec/user on NVIDIA GB200 NVL72. Performance will further improve with future optimizations like Dynamo, NVFP4, and CUDA kernels.