Does it land inside the window?
Finality is a distribution, not a number. Drag the marker to set how long your agent takes and read the probability it beats the block.
Probability the agent decides before finality
84.0%
S1 Nano in training. Runtime in private beta.
AI models trained on Solana itself.
The training and inference layer for Solana. Models that learned from real transactions, running beside a validator, so your agent can act before the block closes.1
A soliton and a dispersive wave packet under the same transport. One holds its shape. The other spreads until there is nothing left to act on.
Finality target150ms
Slot time200ms
Corpus indexed to slot459,653,000
123
Your agent is the slow part.
Solana settles in about a sixth of a second. A general-purpose model wrapped in tools needs four and a half. Watch the bar fill in real time, because that is how long your agent takes to answer.
Both bars run against the wall clock. Slot ticks are every 200 ms, the real slot time after SIMD-0525.
123
Nobody trained the model.
Every Solana agent framework takes a model that learned from the internet and hands it a toolbox. We start lower down, at the data: real transactions paired with what they actually did on chain.
Corpus atlas. 44,800 transaction embeddings, projected. Drag to turn it.
123
Four layers, and we build all four.
The data, the training loop, the models, and the machines they run on. Each one exists because the layer under it was not enough on its own.
Work climbs from Corpus to Runtime. Each plane is a layer we own rather than buy.
There is no model yet.
Corpus is still being built. Anneal is running on a test cluster. S1 Nano is in training. Every speed and accuracy figure on this site is a target we are holding ourselves to, and it says so wherever it appears.
Open weights. Run the test yourself.
When S1 ships, the weights and the benchmark ship with it, so you can check our numbers instead of taking them from a chart we drew.
- Alpenglow replaced TowerBFT on Solana mainnet-beta on 21 October 2026 at slot 459,653,000. 150 ms is the design target, not a guarantee. Early mainnet readings have ranged from 202 ms to a 404 ms median over 24 hours. solana.com, read 29 October 2026.
- The 4,698 ms round trip is built from published latencies for a frontier model and a public RPC endpoint. The 123 ms figure is a specification for a system that does not exist yet. What is not done.