The next language checkpoint is in training
Call /v1/models today and two entries will answer a request: ogi-language-1, and ogi-language-1-distill for validators on the Light tier. A third is in training. When it promotes, it will be the largest checkpoint this network has served.
That is the announcement, and it is deliberately a small one. The larger one comes when there is a checkpoint hash to put under it.
What "largest" is allowed to mean
One number in the profile cannot move. The language surface fixes the hidden dimension at 4,096 because it matches the shared latent, and that latent is the interface every other surface reads and writes. Widening it would mean re-cutting vision, memory and embodiment to fit a model that does not exist yet. So whatever grows here, it is not width.
The rest of the profile — depth, the sparse-expert count, how often the routed path is taken instead of the dense one — is a per-run decision, and it is published with the checkpoint rather than ahead of it. There is no parameter count in this post because the weights are still moving.
Why now
Pretraining sizes the base model at the compute-optimal point for the training budget available in that phase, then deliberately undershoots it by 20–40%, because the model is served continuously and inference cost dominates training cost over a deployment lifetime. Both halves of that sentence are constraints. Only one of them moved.
The budget is not a decision anyone makes by announcement. It is treasury revenue, which comes from swap fees on the protocol-owned liquidity position and from nowhere else. There is no mint and no emission schedule, so a larger run only becomes affordable once the metabolism has paid for it. A full pretraining run is also the most network-intensive computation in the system, reserved for validators in the Cluster tier, the top hardware tier — so the run is gated on that capacity existing as well.
What is not true yet
Nothing about this checkpoint is live. It is not in the catalogue, it is not on canary routing, and it has no id you can send in a request. Treat this post as intent, not inventory; the same rule applies here as to every dark surface in the catalogue, which is that a thing is not announced as existing before it does.
Promotion is the event worth waiting for, and it is mechanical rather than editorial. Validators run the frozen lang-eval-v1 suite against the candidate, scores are cross-validated across validators and outliers flagged, and a regression against the active checkpoint triggers automatic rollback instead of a discussion. The checkpoint that survives carries a content hash of its weights, a manifest hash of its training data, and a hash chain back to the start of the run — lineage any third party can re-derive.
When those hashes exist, the announcement will be a post here with them in it. Until then, this is a model in training, and the honest version of that sentence is exactly this short.