Abliterated DeepSeek V4.1 Flash cuts refusals from 97% to 4%
mrfakename0 · reddit · 2026-10-18
A day after DeepSeek V4.1 Flash 1017's release, mrfakename0 published GGUF quants of both the base and an abliterated version on Hugging Face.
- Refusals drop from 97/100 prompts to 4/100; in think mode from 98/100 to 9/100.
- Key technique: chat and think modes refuse along different directions, so instead of ablating one refusal direction, the author measured and removed the rank-2 subspace they span.
- Deliberately optimized to keep the model intact rather than remove every last refusal, after several iterations.
- A LoRA is available so users needn't redownload weights; imatrix quants are in progress.
Related event: DeepSeek V4.1 Flash abliterated version cuts refusal rate from 97% to 4%(3 posts)→
More from Models
- Claude Opus dissects agent music synthesis: chroma analysis and spectrograms as stand-ins for hearing — teortaxesTex · 2026-10-18
- Perplexity Decider Flash matches Wald accuracy at nearly 2x speed on M4 Mac — antoine_chaffin · 2026-10-18
- Google team touts progress on Gemini's knowledge work capabilities, citing legal reasoning — aran_nayebi · 2026-10-18
- Gemini tells user it has no web search tool — at Google, the search company — jobergum · 2026-10-18
- DeepInfra cuts DeepSeek V4.1 Flash price 51% as agent demand surges — gharik · 2026-10-18
- SemiAnalysis: vLLM hits 1.3B tokens per $ TCO, 4x better than B300 aggregated serving — woosuk_k · 2026-10-18