Qwen 3.6 27B Loops After 100k Context – Server Config Issue?
Examine Qwen 3.6 27B looping after 100k context; adjust server config or model parameters.
Check for known Qwen 3.6 27B looping issues; reduce context size or adjust llama‑server flags; switch to a newer model if needed.
Summary
Users report that Qwen 3.6 27B enters an infinite loop when the context exceeds 100 k tokens, a problem that persists even after attempts to restart or reset the model. The issue surfaced after switching from Gemma 31B Q5 to Qwen 27B Q8, where the model can code, document, and run tests but then stalls. The command used is llama‑server with a 200 k context limit, 24 context checkpoints, and ngram‑mod spec‑type, among other flags. Despite these settings, the model repeatedly loops, indicating a potential bug in the GGUF or server configuration.
No official fix has been released, and the problem appears linked to the high‑context size rather than the model weights. Users have shared screenshots of the looping behavior and attempted various workarounds, none of which resolved the issue. The community is awaiting a patch or guidance from the Qwen maintainers. In the meantime, reducing context size or adjusting server flags may mitigate the problem.
Key changes
- Qwen 3.6 27B loops when context exceeds 100 k tokens
- Loop persists after restart or reset attempts
- User uses llama‑server with -c 200000, --ctx-checkpoints 24, --checkpoint-every-n-tokens 8192, --spec-type ngram-mod
- No official fix yet; issue likely tied to high‑context size or GGUF
- Community has shared screenshots and attempted workarounds without success