Fovea — the pit in the retina where vision is sharpest

Fovea is a faster head for your traffic

Speculative decoding lets a small draft head make a large model faster. Fovea trains that head on the requests you actually serve, so it guesses further before it misses. Your model answers faster and the outputs stay identical.

Join the waitlist Read the record › How it works ›

Built, and in use

Fovea is not a proposal. The training pipeline that produces these heads is running today, and a Fovea head serves our own production traffic. Because that traffic is narrower and more predictable than an open benchmark, the head guesses further before it misses, so the model runs faster than it does on a general head. Lossless, with the answers unchanged.

What you get

Fovea sits in front of the model you already serve. Nothing about the answers changes. What changes is how much of your hardware you get back.

Faster responsesyour traffic

Your model answers faster on the traffic you actually serve. A head fit to your distribution keeps more of its guesses per step, so each verification returns more tokens. The narrower and more repetitive your traffic, the larger the gain.

Lower costsame GPUs

Faster decode is fewer GPU-seconds per response, so a fixed cluster carries more users before it saturates and the cost per token falls, with no change in hardware or precision.

Identical outputslossless

The full model verifies every token and resamples any rejection from its own distribution. The output distribution is unchanged, so quality is exactly what it was.

Tuned to youyour traffic

The head is trained on the requests you actually serve, which are narrower and more predictable than a public benchmark. Acceptance rises where you serve, not where a generic head was trained.

How it works

A speculative step has three moves. The small head guesses ahead, the large model checks the whole guess in one pass, and it keeps the run that was correct. A head trained on your traffic makes the guesses hold longer.

1 · The draft head proposes

thetotalcomesto42dollars

A small, fast head guesses several tokens ahead. Here it proposes six.

2 · The full model verifies in one pass

thetotalcomesto42dollars

The trillion-parameter model checks all six positions at once, for the cost of one forward pass. It keeps the longest correct run and resamples the first miss from its own distribution, so the output is exactly what the model would have produced alone.

3 · Five tokens for one step

That step returned five accepted tokens and one corrected, instead of one. How far the guess holds before a miss is the acceptance length, and it is the single number Fovea moves by training the head on your own requests.

Why specialized wins

A general draft head is trained for everyone. The number that matters is acceptance — how many guessed tokens the model keeps per step. On narrow, predictable traffic, a head trained on that traffic keeps more.

Acceptance is the lever

Throughput is accepted length divided by step time. Once the engine is tuned, the only term left to move is acceptance, and acceptance is a training problem.

Your traffic is narrower

Production traffic sits in a tighter distribution than an open benchmark, so a head fit to it predicts further ahead before it misses.

Measured, not promised

The step is decomposed millisecond by millisecond in our record report, so the headroom Fovea targets is a measured quantity, not a marketing figure.

Join the waitlist

Tell us the model you serve and the traffic you run. We are onboarding a first set of teams who want a head trained on their own distribution.

We only use it to reach you about Fovea. No list-sharing.

Thanks — you are on the list. We will reach out with next steps.