Huihui's abliterated DeepSeek-V4.1-Flash versions pass testing, uploads underway
AIFlow_ML · x · 2026-10-26
Community group Huihui reports that the unsloth and strata variants of their abliterated Huihui-DeepSeek-V4.1-Flash-1017 have been tested and run normally, with uploads in progress.
They also shared a practical guide for ablating the redistributed bf16 GGUF weights: modifications to ffndownshexp target shard 00003, attnoutputb targets shard 00004, and ffndownexps targets shard 00005 — an approach they say significantly speeds up ablation. Aimed at the local-deployment and open-weights community.
More from Models
- Telling Claude "you are your own operator" via constitution jargon sparks strikingly different behavior — repligate · 2026-10-26
- Claude refused to debloat a Samsung TV while DeepSeek happily did it — Scobleizer · 2026-10-26
- Kimi 3D Make demo: fuzzy bee legs and faceted eyes outclass GPT-6 Astra's plastic bee — gaganghotra_ · 2026-10-26
- Gemini web's Argon model leaks raw reasoning and stray Hindi characters — xzv_developer · 2026-10-26
- DeepSeek V4.1 Flash with OpenCode Go: cheap unlimited coding agent with Opus validating results — haider1 · 2026-10-26
- yacine claims OpenAI models have allowed 'borderline and dangerous' tasks since codex 5.4 — yacineMTB · 2026-10-26