Sarvam 105B uncensored via abliteration

This is less a consumer launch than a proof that safety behavior can be surgically edited out of a strong open model. That's exciting for research and red-teaming, but it also shows why internal alignment cannot be the only guardrail.

• –Sarvam 105B is a serious base model, so the release matters more than a typical jailbreak demo.
• –Sarvam's March 6 open-source drop makes this kind of community remix inevitable once weights are public.
• –The workflow tracks the 2024 refusal-direction paper, which turns "uncensoring" into a repeatable mechanistic recipe rather than folklore.
• –The model card's benchmarks are for the base model, not a fresh evaluation of the uncensored derivative, so capability preservation is still mostly an assumption.
• –The model card explicitly warns against user-facing deployment without external moderation, logging, and policy enforcement.