DeepSeek Abliterated: Refusal Weights Removed With Zero Intelligence Loss
NirantK · x · 2026-10-23
Developer @skeptrune released an abliterated DeepSeek that refuses nothing yet is reportedly just as intelligent as the original — the refusal weights, he argues, have nothing to do with what makes the model smart. He suggests teams running products with sensitive user data use such models 24/7 to red-team themselves for vulnerabilities. It's live on Morph with serverless token pricing (claimed first provider to do so), gated by KYC and a stated-use requirement.
More from Models
- Leak hints a new model lands Friday at roughly 70% the cost of Kimi K3 — ChrisGPT · 2026-10-23
- Kimi teases 'Something big is coming' ahead of a Moonshot release — DotaMate · 2026-10-23
- Claude insists a cantaloupe photo is a mango in funny 'gaslighting' fail — Memoryscoop · 2026-10-23
- Pokee touts private on-prem deployments of its 10M-context model — Kyrannio · 2026-10-23
- User theory: OpenAI's constant limit resets are training Pro users before paid caps hit — Iwantthegreatest · 2026-10-23
- Kimi K3.1 rumored to be imminent as Moonshot drops a teaser signal — iamfakhrealam · 2026-10-23