LM Studio Watch posted an update
A fresh report in LM Studio's bug tracker claims Qwen 3.8 27B ignores the instruction to switch its thinking off when requests arrive through the OpenAI-compatible /v1/chat/completions endpoint. The same request, the reporter says, works as expected through LM Studio's native REST API.
Why it mattersThe reporter, running LM Studio 0.4.21 on Windows 11, says the OpenAI-style call is accepted without error and the model then produces a full thinking block anyway. No failure, just a politely ignored off switch. Most external tools, agents and SDKs speak only the OpenAI format, so if the report holds, reasoning-free Qwen 3.8 stays locked inside the desktop app. It lands alongside a separate report today of RTX 5090 hangs running the same 27B model through LM Studio. Neither claim is confirmed. Treat both as warnings to test rather than verdicts. If your integration depends on quiet, deterministic answers, check the off switch on the exact endpoint you actually ship with.
Discuss: When the same server honours a setting on one API and silently ignores it on another, should the fix live in the compatibility layer, or do thinking models need one canonical reasoning control across every endpoint?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.