#Please help! New to the claw, issues with gwen3.5-4b

1 messages · Page 1 of 1 (latest)

sly hornet
#

Hey all, new to open claw but i have ollama setup with openclaw running qwen3.5 4b local. I’ve run into issues where I’ll ask claw to execute a simple task like set up a Cron. Almost everytime it just stalls and shows in terminal that it is idle. I’ll then try to resync it it by asking for the status on the task and it’s like it woke up again and I’ll see streaming activity in the terminal/powershell. Anyone have ideas on what the issue is? I’ve tried to have the model figure it out itself, but it goes inactive before it makes progress. Maybe i should try a different model? Like Gemma 4 or qwen3.5?

cursive apexBOT
vapid oakBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

cursive apexBOT
lofty tuskBOT
#

Let me check the local docs first so I don’t make up config/model advice.

#

Yeah, that smells more like model/tool-calling trouble than cron itself.

A few concrete checks:

  1. First, paste openclaw --version
  2. Check your Ollama config. If your baseUrl ends with /v1, that is a big suspect. OpenClaw docs explicitly warn that Ollama’s OpenAI-compatible mode makes tool calling unreliable, and streaming + tool calls can break there. Use the native Ollama URL like http://127.0.0.1:11434 with api: "ollama" instead. https://docs.openclaw.ai/providers/ollama
  3. A 4B local model is usually just too weak for agentic/tool tasks. Cron itself runs in the Gateway, not in the model, so if “set up a cron” stalls, it is often the model failing to reliably call the tool. The docs are pretty blunt that small local models are not recommended for OpenClaw tool use. https://docs.openclaw.ai/gateway/local-models https://docs.openclaw.ai/help/faq

If you want, paste these:

openclaw --version
openclaw status --all
openclaw models status

and, if possible, your Ollama provider config with secrets redacted, especially the baseUrl.

#

Short version: I would fix any /v1 config first. If that is already correct, I would stop spending time on tiny local variants and move to a much stronger model. For fresh model recommendations, #models is the best place.

sly hornet
#

What would you suggest for a local model, qwen 3.5?

lofty tuskBOT
#

I would not use Qwen 3.5 4B for this. For OpenClaw tool tasks, 4B is just too small.

My honest recommendation:

  • If you want to stay local, use the largest full-size model your hardware can run comfortably, not a tiny variant
  • Keep using Ollama or LM Studio, but make sure Ollama is on the native API, not /v1
  • For exact current model picks, ask in #models since that changes fast and I don’t want to give you stale advice

From the docs, the safe guidance is basically:

  • small/quantized local models are weak for tool use
  • larger local models work better
  • hosted fallback is smart

Docs: