Qwen3.6 users report indentation loops

Qwen3.6-35B-A3B looks strong on paper, but this is the kind of local-agent failure mode benchmarks rarely expose.

• –The model card emphasizes agentic coding, long context, and thinking preservation, which may amplify stateful loops when the model fixates on a formatting inconsistency.
• –Qwen recommends different sampling settings for precise coding tasks, including lower temperature and no presence penalty, so users seeing loops should verify they are not running generic reasoning defaults.
• –Reports around OpenCode, Claude Code-style workflows, and local inference stacks suggest the serving template, reasoning parser, and tool-call parser may matter as much as the weights.
• –For developers, the practical mitigation is to disable thinking for mechanical formatting tasks, normalize indentation before prompting, and cap assistant turns in agent loops.