DeepSeek V4.1 Flash ships a two-stage refusal guardrail, uncensoring effort finds
cephaloform · x · 2026-10-20
Japanese developer @Hikari07jp discovered while abliteration-tuning DeepSeek V4.1 Flash 1017 that it embeds a two-stage guardrail: the first is a standard refusal, and once removed the model doesn't just comply — it produces a second-stage "that's dangerous, but I can help with this safer alternative" response instead. Most models only have the single refusal layer.
The author notes more edit sites degrade model quality more, making a high-quality abliterated version harder than for Ling 3.1 Flash, which was trivially easy to uncensor.
More from Models
- User swaps Copilot 365 sessions to Claude Opus 5.5 and stuck sessions vanish — koltregaskes · 2026-10-20
- Stanford Zou Lab launches Collaborative Intelligence Leaderboard; OpenAI models rank best teammates — james_y_zou · 2026-10-20
- Researcher criticizes missing ablations in new model tech report, citing Llama precedent — eliebakouch · 2026-10-20
- Benchmaxxing score flags models ranking higher on public benchmarks than held-out ones — airesearch12 · 2026-10-20
- DeepSeek model catches its own backup mistake, says "oops," and invents a lateral fix — EffectiveMedium2683 · 2026-10-20
- Exa 3.0 ties Perplexity at #1 on Artificial Analysis Search Index with score of 80 — yoimnotkesku · 2026-10-20