Manta: everything you give it, it keeps

Today we are opening the waitlist for Manta, our long-context model. Manta reads a hundred million tokens in one prompt: a large codebase with its whole history, every filing of a sector, years of a project's documents. What you give it, it keeps.

You can join the waitlist now, and watch the film on our homepage.

What we tested

A long context only helps if the model can find what is in it. So we gave Manta every S&P 500 filing, 88 million tokens, hid 20 codes throughout them and asked for the codes back. It found 18.

The last test makes it harder on purpose: the codes look alike, so a model that only roughly remembers them gets them wrong.

These are recall tests. They show that Manta finds a specific piece of information in a very long input. They don't yet show how well it reasons over all of it, and we will publish more results with the public release.

Using it

Manta speaks the OpenAI API, so the code and tools you already have work with it. You send it everything you have in one request instead of cutting it into pieces and searching through them first.

Early access

We will start with a small group: researchers, engineers and teams with work that needs a model to keep a lot in mind at once. When you sign up, tell us who you are and what you would use Manta for. A person on our team reads every signup.

One condition comes with early access: until the public release, please don't share Manta's outputs or details about it without asking us first. You are welcome to tell people what you think of it.

What comes next

Next we are working toward a billion tokens in one prompt. If you have work that needs that much, we would like to hear about it at empero.org/manta.