OpenClaw + Ollama LLM Timeout: What's Causing It and How to Fix It

In your openclaw.json, add or update the timeout setting for your local model provider:

{
  "models": {
    "local": {
      "provider": "ollama",
      "model": "qwen3.5:9b",
      "requestTimeoutMs": 300000,
      "streamTimeoutMs": 600000
    }
  }
}

requestTimeoutMs: 300000 = 5 minutes. streamTimeoutMs: 600000 = 10 minutes. Adjust based on your hardware. CPU-only with a 30B+ model may need even higher values.

By default Ollama unloads models from memory after 5 minutes of inactivity. When OpenClaw's heartbeat fires 30 minutes later, Ollama has to reload the model — adding a 30–90 second cold-start delay before the first token. That alone can trigger the timeout.

Set Ollama's keep-alive to indefinite by adding an environment variable to your Ollama systemd service or startup command:

# For systemd: edit /etc/systemd/system/ollama.service
[Service]
Environment="OLLAMA_KEEP_ALIVE=-1"

# Then reload:
sudo systemctl daemon-reload && sudo systemctl restart ollama

OLLAMA_KEEP_ALIVE=-1 keeps the model loaded in memory until Ollama exits. Only do this if you have enough RAM. If memory is tight, use OLLAMA_KEEP_ALIVE=30m to extend to 30 minutes instead.

If you're running a 31B model on a machine without a GPU, you're asking for timeout trouble. CPU inference on a large model is simply too slow for OpenClaw's conversational use case.

Practical hardware-to-model guidance:

• 8–16GB RAM, CPU-only: qwen2.5:7b Q4, qwen3.5:9b Q4, gemma3:9b Q4 — these generate at 8–15 tokens/sec on modern CPUs
• 16–32GB RAM, CPU-only: qwen2.5:14b Q4, qwen3:14b Q4 — 4–8 tokens/sec, works but slow
• GPU available (4–8GB VRAM): llama3.3:8b, gemma3:12b — fast enough for any timeout setting
• 31B+ models on CPU: Not recommended for interactive use. Fine for overnight batch tasks, not for heartbeat-driven conversations.

# Pull a faster model
ollama pull qwen3.5:9b

# Update openclaw.json
"model": "qwen3.5:9b"

OpenClaw supports multiple model profiles. Route fast/simple tasks (heartbeat checks, short responses, memory reads) to your local model and complex reasoning to a cloud API. This cuts your API cost by 70–90% while keeping response quality and speed high where it matters.

{
  "models": {
    "default": "local",
    "local": {
      "provider": "ollama",
      "model": "qwen3.5:9b",
      "requestTimeoutMs": 300000
    },
    "powerful": {
      "provider": "anthropic",
      "model": "claude-sonnet-4-6"
    }
  }
}

In your SOUL.md or AGENTS.md, instruct the agent to use model: powerful only for complex multi-step tasks. Heartbeats and simple replies stay on local at $0.