The runtime for AI,
distributed by design.
ORVEX routes every request through a global mesh of workers. Inference happens where it is cheapest, fastest and closest settled invisibly on-chain, delivered like software.
Not a model. Not a chain. A runtime.
Models come and go. Chains settle value. In between lives the layer that actually runs the request: routes it, streams it, bills it, retries it. That layer has always belonged to one provider at a time. ORVEX rebuilds it as an open network.
Every request, best path
Latency, price, availability and locality are evaluated on every call. The router picks the worker most likely to answer well, right now.
Workers you don't manage
Independent operators run inference on hardware they own. You get a single endpoint. They get paid per token served.
Blockchain, invisible
Solana handles settlement, quotas and receipts. You never see a wallet. Payment is in credits. On-chain is an implementation detail.
Ships like software. Runs like infrastructure.
One SDK. One key. Streaming from the first token. The network beneath it is elastic, distributed and priced per unit of work not per seat, not per model, not per month.
One line, any language
$ npm i @orvex/sdk $ orvex login
Familiar shape, better runtime
// stream tokens from anywhere const res = await orvex.chat({ model: "orvex-1", stream: true, messages, });
Experience ORVEX Runtime.
A real chat streaming against the same architecture you see above routed, executed and settled on the mesh in real time. This is what your users will feel.
Same client, same schema. Point your SDK at orbit nothing else changes.
Tokens flow from the first millisecond. No polling, no wait for the full body.
Multimodal on the same endpoint. Send images alongside text with one call.
214 workers across 12 regions. Requests land on the closest healthy node.
Prompts and completions are never persisted. Zero retention by default.
The router scores every path per call. You always get the fastest live worker.
The infrastructure is the artwork.
No stock diagrams. No brains, no hexagons, no globes. Just the actual shape of the system nodes, edges, queues, streams. Alive on every page.
Move your first request in an afternoon.
A drop-in SDK, an OpenAI-compatible endpoint, and a network that scales the moment you send traffic. No infra to provision. No chain to touch. Just a runtime.