OpenClaw, Ollama Telegram bots bog down

The local-first pitch is solid, but this thread shows the operational tax: getting an agent to answer in a chat app is often harder than getting the model to run.

• –OpenClaw’s docs explicitly support Telegram plus local models through Ollama, so this is a real first-class use case, not a weird edge case.
• –The bottleneck may not be raw model speed alone; related community reports point to streaming, polling, token handling, and gateway config as common failure points.
• –`qwen3:8b-nothink` being “free and local” does not make it cheap in latency terms, especially if the model is underpowered for the agent loop.
• –For developers, the lesson is that agent UX depends as much on transport and orchestration as on the model.
• –This is exactly the kind of setup where a small config tweak can matter more than swapping to a bigger model.