NVIDIA Beats Everybody To DeepSeek V4 With Day-0 Blackwell Help, Pushing 3,500 Tokens Per Second On 1.6T Fashions

DeepSeek V4 is out, bringing main optimizations, together with as much as 1.6T mannequin sizes, and NVIDIA is prepared with Day-0 help on Blackwell GPUs utilizing NVFP4. NVIDIA Blackwell NVFP4 Structure Delivers Main Pace-Ups In DeepSeek v4 With Extra Optimizations On The Method With the launch of DeepSeek V4, we noticed some main optimizations in compute & reminiscence necessities. The up to date AI mannequin makes use of simply 27% of single-token inference FLOPs & 10% of the KV cache when working a one-million-token context window. Two new fashions had been additionally launched, one being a Professional mannequin with a parameter measurement of 1.6T, and a Flash model…