High throughput
The rate limits are in the docs, not up to chance
Our own compute sets the ceiling, and the quotas are published.
Throughput starts with capacity, and capacity starts with the cards actually being yours.
"Unlimited" costs nothing to claim. What actually determines your throughput is three things: how much compute the supplier controls, whether the limits are published, and what happens at capacity. A relay reselling somebody else’s quota controls none of them — its ceiling is granted, its rules can change, and the cheapest thing to do under load is hand your request to a cheaper model. Our capacity comes from clusters we own, which is why all three can be written down for you.
How it actually works
- The capacity is ours
- Open weights run in facilities we build out and lease directly, so scaling is a purchase decision we make rather than quota we wait for. It also means nobody else’s ban takes the service down with it.
- Rate limits are published
- Per-key RPM, TPM and concurrency ceilings are in the docs as commitments, not as indications. Need more? Ask, and we answer against real capacity instead of agreeing first and throttling later.
- A 429 at capacity, never a silent queue
- When capacity is reached you get a standard 429 with Retry-After and you decide whether to back off, queue or switch models. Hanging your request indefinitely and quietly rerouting it to something cheaper are the two things we refuse to do.
- Streaming first, and no lying about latency
- Every model streams, so output starts as soon as the first token exists. To be clear though: time-to-first-token rises with input length on very long contexts. That is how attention works, and anyone claiming "a million tokens with zero latency" is not being straight with you.
Check it yourself
None of these four steps need us. You can run them right now.
- 1
The limits in the docs are commitments. Push to them and see whether they hold — if they do not, send us the request ID.
- 2
Measure time-to-first-token with your real input lengths, not an average. A 1K prompt and a 200K prompt are not the same request.
- 3
Deliberately trigger a 429 and check that the response carries Retry-After and that a retry recovers cleanly.
- 4
The status page keeps historical availability. Don’t just look at today — scroll back a few months.
The follow-up questions
Why not just say "unlimited"?
Because it cannot be true. Every service has a ceiling; the only question is whether it is written down or hidden. The hidden kind is the one you discover on your highest-traffic day.
Can concurrency limits be raised?
Yes. Tell us your expected RPM/TPM and traffic shape and we will assess against real capacity and come back with an answer and a date. When we cannot do it we say so, rather than agreeing and quietly throttling.
Do you control the rate for proprietary models too?
Not entirely. Proprietary models come through the vendor’s official commercial channel, so the ceiling is ultimately bound by upstream quota. What we can do is allocate the quota we hold transparently and write it down in the docs. Open models run on our own cards, and that capacity is fully ours.
The other commitments
Don’t trust this page — test it
Take the trial credit and work through the verification steps above one by one. Anything that does not line up, send us the request ID.