When the model “thinks,” the editor sees blank

What happened

With thinking on, Gemma (and later Qwen) put answers in reasoning / reasoning_content while OpenAI-compatible content stayed empty.

Fix pattern

• Default coding aliases to think off / reasoning_effort: none
• Optional *-think twins when you actually want the chain
• Raise client timeouts so a cold 27B load isn’t misread as a dead connection

Takeaway

If the chat pane is blank, check whether the model is thinking into a field your client doesn’t show.

← All Builds

Sister site: toklanes.com — measured tok/s review-bench.