Abliterated DeepSeek refuses nothing with no intelligence loss, provider requires KYC
teortaxesTex · x · 2026-10-23
- Developer @skeptrune announced an "abliterated" DeepSeek: the weights responsible for refusals were deleted, and the model reportedly refuses nothing while remaining as intelligent as the original.
- His argument: refusal behavior is unrelated to intelligence, and teams building products with sensitive user data should run an abliterated model against themselves 24/7 to probe for jailbreak vulnerabilities.
- It's live on Morph with serverless token pricing, billed as the world's first such provider — but requires KYC and an explanation of intended use. A rebuilder mocked the "first in the world" claim given abliteration techniques already exist.
More from Models
- Dev complains one small task burned 30% of Claude Pro quota, mulling $200 tier — lxfater · 2026-10-24
- Leaked Claude UI screenshot shows unannounced Opus 5.5, Opus 5 and Fable 5.1 models — adonis_singh · 2026-10-24
- Developer on Claude Opus 5.5: "nothing else even comes close right now" — prasenx · 2026-10-24
- NYU/UT Austin paper: LLMs write convoluted text due to length-penalty optimization — menhguin · 2026-10-24
- 50 pre-ChatGPT novels tested against Pangram AI detector: zero false positives — TuhinChakr · 2026-10-24
- Seven independent tests find AI detector Pangram false-positive rate near zero — TuhinChakr · 2026-10-24