I abliterated Sarvam-30B and 105B - India's first multilingual MoE reasoning models - and found something interesting along the way!
Reasoning models have *2* refusal circuits, not one. The `<think>` block and the final answer can disagree: the model reasons toward compliance in its CoT and then refuses anyway in the response.
Killer finding: one English-computed direction removed refusal in most of the other supported languages (Malayalam, Hindi, Kannada among few). Refusal is pre-linguistic.