Why this, and why now.

On 21 October 2026 Solana replaced TowerBFT with Votor and the finality target fell from 12.8 seconds to 150 milliseconds. Slots had already been cut to 200. The chain got roughly eighty times faster and the models did not move at all.

Every agent framework on Solana today takes a model that learned from the open internet and hands it a toolbox. That was a reasonable design when a block took twelve seconds. It is not one now. Nobody has trained the model, so we are starting at the data and working up.

How we intend to behave.

Written down so it can be held against us later.

Open weights

Every S1 model under Apache 2.0, with the training recipe and the data card. Not a research licence, not weights-available with a field-of-use carve-out.

Publish the control, not just the win

No reinforcement learning result goes out without its random-reward control beside it. Work published in 2026 showed spurious rewards lifting some base models nearly as much as real ones, and that is now table stakes.

Say what is missing

Every target on this site is marked as a target and listed on one page. When a number turns out to be wrong, the page changes rather than quietly dropping it.

A benchmark you can turn on us

Soliton Bench ships with the weights, graded by replay rather than by a judge model, so a result can be reproduced by anyone who doubts it.

No measured savings claims

We will not publish a figure about what Soliton saved anyone until we have run it against a control for long enough to mean it.

What happens next.

No dates, because we would only be guessing at them. The order is firm.

Now. Corpus replay and S1 Nano training

Replay runs backwards from slot 459,653,000. Nano is in training against the fork grader. Runtime is in private beta with a handful of teams.

Next. S1 Nano weights and Soliton Bench

The first weights release, under Apache 2.0, with the evaluation harness in the same repository so the numbers can be checked the day they are published.

After. S1 and S1 Verify

The everyday model and the refusal model. Verify matters more than it looks: most of the damage an agent does is signing something it should have declined.

Not scheduled. S1 Pro, and an adversarial grader

Pro needs the rest of the stack working first. The grader needs to model a searcher reacting to our agent, which it does not do yet, and route scores read optimistic until it does.

Open roles.

Remote, or near the hardware in Frankfurt, Amsterdam, Tokyo. Write to us with something you have built, not a resume.

Training lead

Owns Anneal. Reinforcement learning where the reward comes off a forked validator rather than a judge model. You have shipped an RL run that survived contact with a real environment. Apply

Validator engineer

Owns the fork cluster and the replay harness. You have run a Solana validator under load and know exactly where it hurts. Apply

Data engineer

Owns Corpus. Ledger replay at scale, and an honest public account of the gaps in it. Apply

Systems engineer

Owns Runtime. Inference on the same rack as a validator, signing inside an enclave, and a policy engine nobody can talk their way around. Apply

Get in touch.

One address, read by the people building it. If you run a validator, or you are already losing money to the gap this site is about, that is the most useful mail we can get.