deepseek-v4.1IMAGE

DeepSeek V4.1 Flash EXL3 3.75bpw 量化发布，适配 3 台 DGX Spark

新量化在 3.5bpw 基础上对误差最大的 gate/up 用 4-bit，对 DeepSeek FP8 的 KL 比 3.5bpw 好 4–20%，已上传 Hugging Face。

Thomas Eastham

@easthambuilds9 天前收录原推发布于 2026.10.22

DeepSeek V4.1 Flash EXL3 3.75bpw 量化发布，适配 3 台 DGX Spark

Thomas Eastham @easthambuilds原推发布于 2026.10.22

DeepSeek-V4.1-Flash-1017 EXL3 3.75 bpw for 3x DGX Spark, TensorFold TP3 This improves our 3.5 bpw quant plus routed expert bits: 4-bit down proj everywhere, 4-bit gate/up where 3-bit error was highest. KL vs DeepSeek's FP8: 4-20% better than 3.5 bpw https://huggingface.co/easthambuilds/DeepSeek-V4.1-Flash-1017-EXL3-3.75bpw

x.com/easthambuilds/status/2113278319554621708采集于 2026.10.23 12:08