OpenClaw + Ollama + gemma4:26b is fast in raw Ollama, but first heavy OpenClaw turns are extremely slow or hit idle timeout

Users running OpenClaw with gemma4:26b via Ollama encounter significantly slow or timed-out first turns in a session, even though the model responds quickly when queried directly through raw Ollama. This is because the first request after model load triggers a cold-start lasting 13–46 seconds, and OpenClaw's default LLM request timeout is too short for this scenario, causing a timeout-based fallback (HTTP 408) — often silently rerouting to a cloud model. Additionally, OpenClaw appears to have a hardcoded ~61-second HTTP fetch timeout that is not connected to the agents.defaults.timeoutSeconds configuration setting, meaning even models that respond to direct curl requests in 5–15 seconds can be killed by OpenClaw before a first token is received. A known workaround is to pre-warm the model by sending a keep-alive request to Ollama before starting a session, preventing the cold-start delay on the first OpenClaw turn.

Related

• Those of you that run Openclaw with Ollama Pro, do you need the local ollama to use cloud?
• OpenClaude com Ollama Cloud
• Does ollama cloud pro will generate token faster than free?
• Tried running LLMs locally to save API costs… ended up waiting 13 minutes for ONE response 🤡
• Recommended Model for a 4060ti 8gb and 16gb ram

Source: local-ai