deepseek-v4.1IMAGE

Community prunes DeepSeek V4.1 Flash experts in half to 199GB, losing 1.2 points on Chinese code

Developer cliecy halves each layer's 384 experts to 192 in a ~2-bit 199GB GGUF, self-reporting HumanEval-zh 89.0% vs 90.2% for the original, but 94% looping in long-form writing and near-broken Japanese; a custom patch is needed to load it.

Colby

@dsoch8cArchived 9 days agoOriginally posted 2026.10.23

Community prunes DeepSeek V4.1 Flash experts in half to 199GB, losing 1.2 points on Chinese code

Colby @dsoch8cOriginally posted 2026.10.23

Open original on X

社区开发者 cliecy 将 DeepSeek-V4.1-Flash-1017 专家层砍半并压缩，自测中文代码成绩仅比原版差 1.2 个百分点。北京时间 10 月 23 日，作者将每层 384 个专家减至 192 个，仅保留中文代码常用部分，制成约 2 比特、199GB 的模型文件，目标是装进两张显存 48G 的显卡。作者自测且无第三方复现显示，HumanEval-zh 中精简版得分 89.0%，原版 90.2%；同时作者自测其长文写作循环率 94%，日文几乎报废。该模型无法直接使用，读者若想体验，必须安装作者提供的专用工具补丁才能加载。 https://huggingface.co/cliecy/DeepSeek-V4.1-Flash-1017-V6-codezh-reap192-IQ2XS-GGUF

x.com/dsoch8c/status/2113468859412066310Archived 2026.10.23 12:06