deepseek-v4.1IMAGE

ISTA releases compressed DeepSeek V4.1 Flash 1017 GGUF claiming 100% recovery

Austria's ISTA lab published a non-uniform compressed build of the 763.2B model; the 248.84GB variant averages 93.54 with a claimed 100% recovery, but requires the Strata engine as llama.cpp lacks support.

Colby

@dsoch8cArchived 8 days agoOriginally posted 2026.10.24

ISTA releases compressed DeepSeek V4.1 Flash 1017 GGUF claiming 100% recovery (image 1 of 3)

ISTA releases compressed DeepSeek V4.1 Flash 1017 GGUF claiming 100% recovery (image 2 of 3)

ISTA releases compressed DeepSeek V4.1 Flash 1017 GGUF claiming 100% recovery (image 3 of 3)

Colby @dsoch8cOriginally posted 2026.10.24

Open original on X

奥地利 ISTA 的深度算法与系统实验室发布 DeepSeek-V4.1-Flash-1017 压缩模型文件，作者自测称性能恢复率达 100.0%。北京时间 10 月 24 日，该实验室在开源平台上线了这套 763.2B 参数大模型的非均匀压缩版本。在作者自测的单次跑分中，推荐的 248.84GB 规格三项测试平均 93.54，标称恢复率 100.0%，但第三方尚未复现；另一款 231.32GB 规格恢复率为 98.7%。目前通用的 llama.cpp 工具尚不支持该模型，用户必须搭配专用的 Strata 引擎使用，且官方尚未公布实际出字速度。 https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF

x.com/dsoch8c/status/2113815222050075034Archived 2026.10.24 12:06