Tag
Someone has optimized DeepSeek V4.1 to run faster than the official API on NVIDIA A100 GPUs, despite A100 not supporting FP4 precision, and shared the implementation on GitHub.