three.ws · Decentralized inference

Same prompt, two providers, real responses. The left card hits three.ws's default inference (Claude). The right card routes the identical request through a decentralized AI gateway whose orchestrators run open-weight models on commodity GPUs. Phase 4 of the roadmap: decoupling inference from any single vendor so agents stay online when one provider doesn't.

Prompt

Max tokens Temperature

Settings · decentralized gateway model

Open-weight model

three.ws default

Claude · Anthropic

latency — in — out —

awaiting prompt…

model · —

three.ws decentralized

Open-weight · decentralized

latency — in — out —

awaiting prompt…

model · — · gateway · —

ready.

Powered by Livepeer AI Gateway on the decentralized side. Claude reply provided by Anthropic. The public decentralized gateway is rate-limited and occasionally slow — typical p50 latency is 2–10 s for an 8B-param model.