Hi everyone! this weekend I shipped a quant for the Flash-Base model in the deepseek V4 series. I posted all the quality, throughput and verification metrics in the repo:
https://huggingface.co/EnsueAI/DeepSeek-V4-Flash-Base-INT4
lmk what you think!
It is the full 284B params in 157 GiB at full FP8 speed. I ran most of my tests on 4 H100s with about 320 GB of VRAM.