Pre-launchS1 Nano is in training. Runtime is in private beta.What is not done

S1

Four models, one shared training set.

They read the same transactions during training and share one tokenizer, so they agree on what an address, an instruction and a compute budget are. You pick one by how long you can afford to wait.

Time to a short plan, drawn against the 150 ms finality window. Targets, not measurements.

S1 Nano

Small enough to answer while a block is still being built. Cancel, re-route, hedge or stand down.

Size
4B
Reads at once
32K
Time to answer1
46 ms
Status
In training

S1

The everyday Solana engineer. Reads program interfaces, explains a failed transaction, plans a multi-hop route.

Size
32B
Reads at once
256K
Time to answer1
380 ms
Status
Spec frozen

S1 Pro

Thinks for an hour. Protocol design, audit review, working out how an exploit happened.

Size
140B
Reads at once
1M
Time to answer1
2.1 s
Status
Spec frozen

S1 Verify

Writes nothing. Reads a transaction and its simulated result, then says sign or do not sign.

Size
2B
Reads at once
16K
Time to answer1
14 ms
Status
Private beta

One tokenizer, built for addresses.

A general vocabulary was fitted to English and code. Base58 is neither, so an address shatters and the model spends its attention on spelling. Type into it and watch.

General tokenizer44 tokens
9WzDXwBbmkg8ZTbNMqUxvQRAyrZzDsGYdLVL9zYtAWWM
Soliton tokenizer4 tokens
<addr>9WzDXwBbmkg8ZTbNMqUxvQRAyrZzDsGYdLVL9zYtAWWM
Compression11.0xContext recovered91%Characters44
The upper lane is a byte-pair merge against a vocabulary fitted to English and code, which is what every general-purpose model brings to a chain. Base58 has no word structure for it to find, so an address shatters. The lower lane recognises an address and carries it as a type marker plus three limbs of the decoded key.

Weights

Download it, run it, check it.

Every S1 model is released under Apache 2.0 with the training recipe, the data card and the evaluation harness. If our numbers are wrong, you will be able to show it with the same files we used.

See how they are graded

Nothing is published yet. S1 Nano is in training.

  1. Time to answer is the target for a short plan on colocated hardware, and it is a specification rather than a measurement. No S1 model has published an evaluation run. What is not done.
  2. Sizes and context lengths are frozen design specifications as of 29 October 2026. Tokenizer behaviour on base58 was checked against the cl100k_base vocabulary.