OpenClaw + Remote Ollama Fails: "does not support thinking" — But Direct API Works Fine

Published April 18, 2026 · Bug #67949 · Affects: 2026.4.12+ · Setup: OpenClaw → remote Ollama on LAN

You've set up OpenClaw on one machine and Ollama on another machine on your local network. Direct curl calls to Ollama's /api/chat endpoint work perfectly. But when you run:

openclaw infer model run --model ollama/qwen2.5:0.5b --prompt "hello" --json

You get:

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

No response in the OpenClaw gateway UI either. The model is working — OpenClaw is the problem.

What's Causing It

OpenClaw's Ollama adapter sends a "think": true flag in its requests by default when reasoning/thinking features are enabled globally in your config. Most small Ollama models — including qwen2.5:0.5b, qwen2.5:7b, and other quantized variants — don't support the thinking parameter and reject the request with a 400 error.

The frustrating part: when you call /api/chat directly with "think": false (or omit the field entirely), it works. OpenClaw is injecting a parameter the model can't handle.

Fix 1: Disable Thinking for the Ollama Provider in Config

In your openclaw.json, explicitly disable thinking for your Ollama models:

{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://192.168.1.102:11434",
        "models": [
          {
            "id": "qwen2.5:0.5b",
            "thinkingDefault": "off",
            "params": {
              "think": false
            }
          }
        ]
      }
    }
  }
}

Restart the gateway after saving:

openclaw gateway restart

Fix 2: Use the infer Command with Thinking Disabled

For one-off infer runs, pass thinking off explicitly:

openclaw infer model run \
  --model ollama/qwen2.5:0.5b \
  --prompt "Reply with exactly: hello" \
  --reasoning off \
  --json

Fix 3: Set a Global Default for Small Models

If you're running several small Ollama models that don't support thinking, set a provider-level default in your config:

{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://192.168.1.102:11434",
        "defaultParams": {
          "think": false
        }
      }
    }
  }
}

This applies to all models under the Ollama provider unless overridden per-model.

Fix 4: Switch to a Thinking-Capable Ollama Model

If you actually want thinking/reasoning in your local model, use a model that supports it. Ollama models with native thinking support (as of April 2026):

• qwq:32b — strong reasoning, requires ~20GB RAM
• deepseek-r1:8b — lighter reasoning model
• qwen3:8b — Qwen3 family with thinking support

ollama pull deepseek-r1:8b
# Then update your OpenClaw config to use deepseek-r1:8b

Two-Machine LAN Setup Tips

While you're fixing this, a few things to double-check for a reliable OpenClaw ↔ remote Ollama setup:

• Firewall: Port 11434 must be open on the Ollama machine. On Ubuntu: sudo ufw allow 11434
• Ollama bind address: By default Ollama only listens on localhost. Set OLLAMA_HOST=0.0.0.0 in your Ollama service environment to allow LAN connections.
• OpenClaw baseUrl: Use the LAN IP, not localhost: "baseUrl": "http://192.168.1.102:11434"

Need help setting up a two-machine OpenClaw + Ollama deployment? ClawReady has done this setup dozens of times — we can get you running in under an hour.

← Back to blog