Introduction

ORVEX is a distributed inference runtime. One endpoint routes every request to the best-fitting worker, streams tokens back, and settles usage on a signed ledger.

What is ORVEX

ORVEX exposes an OpenAI-compatible API in front of a network of independent GPU workers. You send a chat, embedding, or image request; the router picks the worker with the lowest expected cost that fits your latency and policy constraints; the response streams back over server-sent events; and a signed receipt records what happened.

There is nothing to deploy. Point your existing SDK at https://api.orvex.dev, swap the key, and you are running on the mesh.

214 workers

Global mesh

Requests routed across 12 regions. Cold starts are absorbed by warm pools; p50 first-token is under 40 ms in-region.

drop-in

OpenAI-compatible

/v1/chat/completions, /v1/embeddings, /v1/images. Same request shape, same streaming envelope.

auditable

Signed receipts

Every request produces a hash-linked ledger entry. Verify off the hot path; never block a response on settlement.

per-token

Predictable cost

Prices published per model, quoted before dispatch, charged on completion. No idle GPU time, no minimum spend.

Architecture

The runtime is four layers: the edge gateway that terminates TLS and auth, the router that chooses a worker, the worker mesh that runs the model, and the observer that records the receipt.

Principles

One endpoint. Model selection is a parameter, not a URL.
Streaming by default. The first token is the product; buffering is opt-in.
Prove, don't promise. Every response links to a receipt you can verify without us.
Fail loudly. One error envelope, always.

Where to next

90 seconds

Quickstart

Install the SDK, set a key, stream a completion.

reference

Chat API

Full parameter reference for chat completions.

Quickstart