deepseek-v4.1POST

3 台 DGX Spark 上 EXL3 3.75bpw 跑 V4.1 Flash：4M 上下文、无 OOM

用户分享在 3 台 DGX Spark 上用 TensorFold TP3 以 EXL3 3.75bpw 量化运行 DeepSeek-V4.1-Flash-1017，接近全权重、4M 上下文稳定无 OOM，作为日常主力，并附完整配方。

Thomas Eastham

@easthambuilds2 天前收录原推发布于 2026.10.27

Thomas Eastham @easthambuilds原推发布于 2026.10.27

@marc_jeffrey18 @MiaAI_lab Works today if you're OK with a quant: I run DeepSeek-V4.1-Flash-1017 as EXL3 3.75 bpw on TensorFold TP3 across 3 Sparks. Full recipe: https://github.com/easthambuilds/DeepSeek-V4.1-Flash-1017-EXL3-3.75bpw-3x-DGX-Sparks-TensorFold Nearly full weights and still gives good context size of 4M. stable without OOM issues. This is my daily driver.

x.com/easthambuilds/status/2114840389928587299采集于 2026.10.27 12:09