OpenClaw + Ollama Small Models: "Does Not Support Thinking" 400 Error

Published April 18, 2026 · Bug #67949 · Affects: 2026.4.12+ · Setup: Remote Ollama / small models

You've set up OpenClaw on one machine pointing to Ollama running on another machine on your local network. You pick a small, fast model like qwen2.5:0.5b, run a test, and get silence — nothing in the dashboard, no response in chat. Check the logs and you'll find:

400 {"error":"\"qwen2.5:0.5b\" does not support thinking"}

Meanwhile, a direct curl to Ollama's /api/chat with the exact same model works perfectly. So what's OpenClaw doing differently?

The Cause

OpenClaw injects a think or thinking parameter into Ollama API requests. For larger models that support extended reasoning (like qwen3 or deepseek-r1), this is intentional. But smaller models — qwen2.5:0.5b, phi3:mini, gemma:2b, and others — don't implement the thinking interface at all. Ollama rejects the request with a 400 before it ever generates a response.

OpenClaw doesn't fall back gracefully — it just drops the response, leaving the UI blank and the gateway silent.

Workarounds

Option 1: Disable Thinking for the Model in Config (Recommended)

Add a model entry in your openclaw.json that explicitly disables thinking for the small model:

{
  "models": {
    "providers": {
      "ollama": {
        "models": [
          {
            "id": "qwen2.5:0.5b",
            "thinkingDefault": "off",
            "params": {
              "think": false
            }
          }
        ]
      }
    }
  }
}

Add a similar entry for each small model you use. After editing, restart the gateway:

openclaw gateway restart

Option 2: Use a Model That Supports Thinking

If your hardware can handle it, switch to a model that supports Ollama's thinking interface. These work without any config changes:

ollama pull qwen3:4b
# Then update your openclaw config to point to qwen3:4b

Option 3: Pin to OpenClaw 2026.4.10

If you need the small model specifically and don't want to fiddle with config, rolling back to 2026.4.10 avoids the thinking injection entirely for Ollama providers:

npm install -g openclaw@2026.4.10
openclaw gateway restart

Option 4: Use openai-compat Format with Ollama

Ollama exposes an OpenAI-compatible API at port 11434. Routing OpenClaw through the openai provider with a custom base URL bypasses the Ollama-specific thinking injection:

{
  "models": {
    "providers": {
      "openai": {
        "baseURL": "http://192.168.1.102:11434/v1",
        "apiKey": "ollama",
        "models": [{ "id": "qwen2.5:0.5b" }]
      }
    }
  }
}

This is especially useful in multi-machine setups where you want fine-grained control over which models get which parameters.

Diagnosing the Problem

# Test directly from OpenClaw machine to Ollama machine
curl http://192.168.1.102:11434/api/chat \
  -d '{"model":"qwen2.5:0.5b","messages":[{"role":"user","content":"hello"}],"think":false}'
# Should return a valid response

# Now test with think:true — this is what OpenClaw was sending:
curl http://192.168.1.102:11434/api/chat \
  -d '{"model":"qwen2.5:0.5b","messages":[{"role":"user","content":"hello"}],"think":true}'
# Returns: 400 {"error":"\"qwen2.5:0.5b\" does not support thinking"}

Multi-Machine Ollama Setup Tips

While you're here — a few things that trip people up in remote Ollama setups:

Setting up a multi-machine Ollama + OpenClaw environment? ClawReady's setup service covers remote Ollama configuration, model selection for your hardware, and cost optimization across local and cloud providers.

← Back to blog