Expert-only IQ3 requant of DeepSeek-V4-Flash-0731: better KLD than UD-IQ3_S, 1.4x decode on a CPU-spill rig

Hey all, tldr / who this helps: you run a mixed multi-GPU box where the experts spill to RAM, and you want to stay in the 3-bit tier instead of dropping to Q2 to make it fit. https://huggingface.co/TacoTakumi/DeepSeek-V4-Flash-0731-GGUF I requantized only the 129 routed expert tensors of DeepSeek-V

中文全文翻译正在补齐，可先查看下方英文原文。

Hey all, tldr / who this helps: you run a mixed multi-GPU box where the experts spill to RAM, and you want to stay in the 3-bit tier instead of dropping to Q2 to make it fit. https://huggingface.co/TacoTakumi/DeepSeek-V4-Flash-0731-GGUF I requantized only the 129 routed expert tensors of DeepSeek-V