Qwen3.6 is maintaining context inside the CoT

Alibaba's Qwen3.6, the latest iteration in its open-source large language model series, is showing a significant advancement in reasoning transparency. User testing reveals that when prompted with a specific command-line argument, the model can maintain and reference the internal numerical values it generates during its Chain-of-Thought (CoT) reasoning process. This means that in a multi-step math or logic problem, if the model deduces a number in step two, it can reliably recall and use that same number in step four, rather than potentially generating a new, inconsistent value. The feature is activated by adding the flag `--chat-template-kwargs '{"preserve_thinking": true}'` to the model's inference command.

This capability addresses a common failure mode in AI reasoning known as "reasoning collapse," where models lose track of their own intermediate conclusions. While testers note it can sometimes be challenging to make the model rigidly adhere to a single format, the core function of remembering self-generated context works reliably. This development moves Qwen closer to performing more deterministic, verifiable reasoning, which is critical for applications in coding, data analysis, and scientific computation where each step must be auditable and consistent.