Fix: Ollama Local Model Stuck Thinking (CPU / No GPU)

CPU inference is orders of magnitude slower than GPU. A 7B model that responds in 3 seconds on a GPU can take 5–10 minutes on a modern CPU — long past OpenClaw's default timeout. The fix is either a smaller model or a longer timeout (or both).

What You See

Fix A: Switch to a Smaller Model

This is the most impactful change. On CPU-only systems, model size directly controls whether the tool is usable at all:

Pull a smaller model and update your config:

ollama pull llama3.2:1b
# or
ollama pull qwen2.5:1.5b

{
  "llm": {
    "provider": "ollama",
    "baseUrl": "http://localhost:11434",
    "model": "llama3.2:1b"
  }
}

Fix B: Increase the Timeout

If you want to keep using a larger model and accept the slower speed, increase requestTimeout to give Ollama enough time to respond:

{
  "llm": {
    "provider": "ollama",
    "baseUrl": "http://localhost:11434",
    "model": "llama3.1:8b",
    "requestTimeout": 600000
  }
}

600000 is 10 minutes. This is a last resort — the UX will be very slow. Combining a smaller model with a 120s timeout is a much better user experience.

Fix C: Verify Ollama Is Actually Running

Before tuning timeouts, confirm Ollama is responsive:

# Is Ollama running?
curl http://localhost:11434/api/tags

# Test a real inference (watch the timing)
time curl http://localhost:11434/api/generate \
  -d '{"model":"llama3.2:1b","prompt":"say hi","stream":false}'

Fix D: Model Name Must Be Exact

If the model name in openclaw.json doesn't match what's pulled in Ollama, the request silently fails. Check the exact name:

ollama list

Use the exact name from the NAME column — including the tag (e.g. llama3.2:1b, not just llama3.2). If the tag is latest, you can omit it or include it — either works.

CPU Performance Expectations