NVIDIA Beats Everyone to DeepSeek V4 With Day-0 Blackwell Support, Pushing 3,500 Tokens Per Second on 1.6T Models

DeepSeek released V4 with major efficiency gains, plus new 1.6T and 284B model variants.

NVIDIA said Blackwell GPUs support DeepSeek V4 on day one, using NVFP4 and FP4-based optimizations.

Early testing on GB300 showed promising inference speed, with throughput near 3,500 tokens per second for the 1.6T model.

The news underscores Blackwell’s readiness for very large, long-context AI workloads and broader future FP4 compatibility.