New todaydeepseek-v4.1POST
Testing MXFP8 Engram weights instead of 4-bit in a DeepSeek V4.1 Flash 1017 IQ2_XXS quant on Strata
The author tries a DeepSeek V4.1 Flash 1017 IQ2_XXS-Uncensored quant on Strata, swapping the SSD-resident Engram part from 4-bit to original MXFP8 weights using only the weights already on hand.
となりのトトノ🏯Local LLM | Tonoken3
@Tono_Ken3Archived 8 hours agoOriginally posted 2026.10.28
となりのトトノ🏯Local LLM | Tonoken3 @Tono_Ken3Originally posted 2026.10.28
Strataで1017での推論精度アップのアイディアを見つけた。IQ2_XXS-Uncensordで、SSDに置くEngram部分を4ビット量子化されたものでなく、オリジナルのMXFP8重みで使ってみるもの 手元にある重みの組み合わせだけで実行できるので試す https://huggingface.co/Starw1/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-IQ2_XXS-MXFP8-Engram-GGUF
x.com/tono_ken3/status/2115239219149095333Archived 2026.10.28 12:01