When running OpenClaw with Ollama locally, there is a measurable difference between the two startup methods.
ollama launch openclaw
Observed behavior:
• Modifies ~/.openclaw/openclaw.json
• Registers multiple models
• Sets agent-oriented defaults
• Uses large context windows (e.g. 40k+)
• Higher RAM usage
Example (qwen3:8b on macOS):
• ollama run qwen3:8b → ~5.5 GB RAM, very fast
• ollama launch openclaw → ~10+ GB RAM, noticeably slower
openclaw gateway start
Observed behavior:
• Uses existing configuration
• Does not auto-register models
• No implicit reconfiguration
• More predictable memory footprint
Context window requirement
OpenClaw requires a minimum context window of 16,000 tokens for Ollama backend.
Practical note
For controlled local setups, start OpenClaw directly and manage models explicitly with ollama pull, instead of relying on ollama launch.