My Kimi is impersonated qwen 3 2.5 3B (local model)
✅Solved
That confirms it — Kimi is still running the cron jobs, not Qwen.
Those log entries are just health check pings (GET /api/ps, HEAD /). An actual Qwen inference would show a POST /v1/chat/completions or POST /api/generate — and there's none during the cron run at 00:50.
So the subagents.model config fix didn't work. The cron runner is ignoring it and falling back to Kimi.
This is likely a bug in OpenClaw worth reporting. In the meantime, the only real workaround would be setting Qwen as the global primary model — but that affects your main session too, which you probably don't want.
Those log entries are just health check pings (GET /api/ps, HEAD /). An actual Qwen inference would show a POST /v1/chat/completions or POST /api/generate — and there's none during the cron run at 00:50.
So the subagents.model config fix didn't work. The cron runner is ignoring it and falling back to Kimi.
This is likely a bug in OpenClaw worth reporting. In the meantime, the only real workaround would be setting Qwen as the global primary model — but that affects your main session too, which you probably don't want.
Solution
On a 24GB M4 Mac mini, the practical “bigger local model” sweet spot is usually:
- ~14B instruct @ Q4 (often the best balance of quality/speed; plenty of headroom)
- ~20–27B instruct @ Q4 (noticeably smarter, but slower; still generally doable on 24GB if you don’t crank context length too high)
- ~32B @ Q4 is sometimes possible but tends to get tight/slow on 24GB once you add KV cache (larger context = a lot more memory). I wouldn’t target this first.
Two important gotchas for your specific problem (cron + tools):
1) Tool-calling reliability isn’t guaranteed even if you go bigger. 3B → 14B helps, but many local instruct models still struggle with “call exactly the right tool with the right JSON schema every time.”
2) For cron jobs, you can often avoid tool-calling entirely: make the job isolated + announce delivery, so the model just outputs text and the Gateway delivers it (no Telegram/Discord tool calls needed). Docs: https://docs.openclaw.ai/automation/cron-jobs
If you tell me what you’re using to host the local model (LM Studio vs Ollama vs llama.cpp) and what context length you set, I can suggest the safest model size + quant + context combo for that stack. Also: for up-to-date “which exact model family is best right now,” the server’s #models channel is the right place (things change weekly), and the general models doc is here: https://docs.openclaw.ai/concepts/models
- ~14B instruct @ Q4 (often the best balance of quality/speed; plenty of headroom)
- ~20–27B instruct @ Q4 (noticeably smarter, but slower; still generally doable on 24GB if you don’t crank context length too high)
- ~32B @ Q4 is sometimes possible but tends to get tight/slow on 24GB once you add KV cache (larger context = a lot more memory). I wouldn’t target this first.
Two important gotchas for your specific problem (cron + tools):
1) Tool-calling reliability isn’t guaranteed even if you go bigger. 3B → 14B helps, but many local instruct models still struggle with “call exactly the right tool with the right JSON schema every time.”
2) For cron jobs, you can often avoid tool-calling entirely: make the job isolated + announce delivery, so the model just outputs text and the Gateway delivers it (no Telegram/Discord tool calls needed). Docs: https://docs.openclaw.ai/automation/cron-jobs
If you tell me what you’re using to host the local model (LM Studio vs Ollama vs llama.cpp) and what context length you set, I can suggest the safest model size + quant + context combo for that stack. Also: for up-to-date “which exact model family is best right now,” the server’s #models channel is the right place (things change weekly), and the general models doc is here: https://docs.openclaw.ai/concepts/models
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,426Members
Resources
Recent Announcements
Similar Threads
Was this page helpful?
