What is Omnious
An RFQ market for LLM inference. Your request goes to auction, providers bid to serve it, and you pay per request in USDC.
Omnious is a request-for-quote market for LLM inference. Companies running open-weight models (Llama, Qwen, DeepSeek, Mistral) stream signed price quotes into the router. When you send a query, the router runs an auction over every live quote, awards your request to the best bid on price and measured latency, and streams the reply back. Payment uses USDC with prepaid credits for production agents: fund once, then a bounded x-credit-key carries authorization on every request. Direct x402 remains available for one-off no-prefund calls.
The current clearing rule is second-score/3, a Vickrey-style reverse scoring auction. It ranks expected request cost after router-measured latency, completion reliability, verification trust, and throughput. The winner clears toward an independent rival's score, subject to a customer cap and own-ask floor; thin books disclose a reserve or own-bid clear. It is not simply “the winner gets the second-best price.”
What that buys you
- A market price, not a list price. Provider marginal costs differ with hardware, utilization, and energy prices. The auction finds the clearing price on every request instead of charging whatever a price list says this month.
- Measured performance. The router streams every reply through its own metering proxy, so latency and token counts are measured facts, not provider claims.
- A receipt you can audit. Every request produces a signed usage record plus a full economic breakdown: the reconstructed auction, what every rival would have charged, and the fee split.
Start with Life of a request to see the whole flow, or jump to how the auction clears.
- router/src/market/auction.ts scoring and second-score clearing
- router/src/http/server.ts the request orchestration