- accept qwen3.5/qwen3_5/qwen35 filenames in the model list (MoE excluded) - skip the lm_head patch for Qwen35Config (no self_attn in layer 0, tied on 2B/4B) - strip completed <think> blocks from replies Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>