/compact works by sending the full conversation to the model with a specific instruction block about how to format the output. the failure with local models is usually not context size but response format - claude was fine-tuned to return compaction output in the format claude code expects, qwen returns something structurally different, cc can't parse it and errors.

easiest debug: watch your local server logs during a /compact attempt. you'll see the exact prompt cc sends including the output format instructions. then try running just that prompt against qwen directly in your api client to see what it actually returns vs what cc expects.

timeout is also worth ruling out separately. local inference on a large context plus the compaction instruction is slow, and cc has request timeouts. if qwen starts generating but cc kills the connection before its done you'd see the same failure with a different root cause.