The fastest AI on the internet.
One endpoint. Every fast model. We race the top inference providers on every request and stream back whichever answers first.
We don't own the silicon. We own the routing.
Three layers we control turn commodity backends into the fastest response on the wire — no custom chips, no lock-in.
Provider racing
Every request fans out to all configured backends at once. The first to emit a real token wins; the losers are aborted the instant we have an answer. Because each provider has bad minutes — a cold pool, a noisy neighbour — racing makes your p99 the best-of-N, which no single vendor can advertise.
first-token-wins · loser-cancelEdge cache
Deterministic requests hit our edge cache and return with zero inference — the cheapest token is the one we never generate. Repeated prompts come back instantly, from the datacenter nearest your user.
0ms on cache hitStreaming passthrough
The winning stream is piped to you byte-for-byte from the closest edge — no re-encode, no added hop. Time-to-first-token is what humans feel, and we shave the network everyone else eats.
edge-local SSEChange one line. Keep your SDK.
OpenAI-compatible. Point any OpenAI client at our base URL and ask for psm-turbo or psm-max.
import OpenAI from "openai"; const client = new OpenAI({ baseURL: "https://fast.purplesquirrel.media/v1", apiKey: process.env.PSM_KEY, }); // same call you already write — just faster const res = await client.chat.completions.create({ model: "psm-turbo", stream: true, messages: [{ role: "user", content: "say hi, fast" }], });
Response headers carry the receipts: x-psm-provider (who won the race), x-psm-tier, x-psm-cache.