Qwen/Qwen3-4B-Base weights with the Qwen3-4B-Thinking-Preservation chat template (thinking always preserved, no nonthinking mode), so the base model can be SFT-trained / evaluated with the same multi-turn thinking contract.

Thinking is always preserved across multi-turn history (append-only). Every assistant turn keeps its <think>...</think> reasoning, not just the latest one, and the generation prompt always opens <think> (passing enable_thinking=False has no effect). This makes multi-turn agent training match evaluation — the model always sees its own prior reasoning. Model weights are identical to Qwen/Qwen3-4B-Base; only the chat template differs.