OpenClaw Chat Latency Slows Local Assistants

The core issue here looks less like a hardware problem and more like using a heavyweight reasoning model on the latency-sensitive front door. If every greeting goes through a 122B-class thinker, the assistant will feel broken no matter how strong the machine is.

• –Split the fast path from the slow path: use a small, responsive model for chat and routing, then escalate to a larger model only when tool use or deep reasoning is needed.
• –The 60-90 second delay strongly suggests prompt size, context buildup, and model choice are dominating runtime, not just raw token throughput.
• –For local-first assistants, UX lives or dies on perceived immediacy; a snappy orchestrator matters more than a brilliant one for casual dialogue.
• –The cloud-main-agent workaround is practical, but it dilutes the privacy and offline advantages that make OpenClaw appealing in the first place.
• –This is a good reminder that agent architectures need tiered inference, not one model trying to do both chat and control-plane work.