Early access beta
DeepSeek V4.1 Flash at up to 1,500 tokens per second.
Token Factory's fastest engine, open to a small group of testers. Point your own tools at an OpenAI-compatible API, or try it in the playground.
Sign-ups are closed for the moment. Testers already in can still sign in.
Up to ~1,500 tok/son code, measured end to end: a 7,000-token game in about 5 seconds. Over 1,100 when two requests share the engine; long prompts and reasoning run about half as fast.
OpenAI-compatible
/v1/chat/completions with streaming. Works with the OpenAI SDKs; tool calls for coding agents are coming.Waitlist and wavesWe let people in in waves. Seats are limited: sessions last 10 minutes when others are waiting, and have no limit when nobody is.
Free during the betaRunning on an NVIDIA GB300 NVL72 rack in our lab. It is a beta: expect maintenance windows.
Signing in shares your GitHub profile and email address with us, nothing else (no repository access). We do not store prompts or answers; we keep per-request numbers (tokens, timing) to tune the engine.
You're on the list
#–
We let testers in in waves. Come back here, or keep the tab open: this page updates by itself.
Playground
0tok/s
Ask anything. Try "Write a complete Tetris game in one HTML file" and watch the counter.
Your API key
No key yet.
Shown once. Store it now; you can always make a new one (the old one stops working).
- Base URL
- Model