Updates from Geocentric.

Our flagship model has 249,396,480 parameters and was trained on one RTX 2060. We are not releasing its weights.

The memory problem that stops a model this size fitting on a 6 GB card, and the change that fixed it.

Writing samples, diagnostic results, and electricity-cost estimates.

A new systems pass measured faster Apple M4 generation and lower optimizer-state memory in controlled tests.

The most capable models Geocentric has trained. Details at release.

Progressive depth, context folding, selective loss, and optimizer-memory work now sit under one training-efficiency system.

An open-weight intelligence and model engineering lab building models, evaluations, and the tooling to run them locally.