GLM-4.7 is the answer

After running several local models for weeks, the author reports that glm-4.7-flash:latest is the best fit for OpenClaw.

Qwen 3.5 27B only drove up GPU usage while actually stalling out often, and Gemma 4 26B repeatedly said "done" even when it hadn't finished the task.

glm-4.7, on the other hand, ran with a 36/64 split even on an RTX 3090 24GB setup, and after the initial cache it kept working continuously, albeit slowly, with no timeouts.

The key impressions were:

• A model that keeps up continuous reasoning rather than quiet execution fits OpenClaw's execution model better
• Once it gets going, responses don't get cut off and it follows through to actually editing files
• During task execution, /abort, /stop, and /cancel felt like they didn't work well

In the comments, people also shared the practical tips that using the native Ollama API can be necessary for tool calling to work properly, and that timeoutSeconds should be increased.