Kestrel
Our largest model. 249,396,480 parameters trained on 5.72 billion tokens of code, Linux, networking and science, on one RTX 2060. Because of security and safety concerns, it will not be released publicly.
Geocentric model family
Meet Kestrel.
Two models: Kestrel, our flagship, which is announced but not publicly available, and Arc, our compact co-writer, which is open source and has a published evaluation.
Arc
A low-cost, experimental English co-writer. 120.8M parameters, a 1,024-token context, and published writing samples alongside its known limitations.
Unreleased model
A further model is in training.
Current small-model work
120.8M parameters. 1,024-token context.
11transformer layers
896model width
32Kvocabulary
1 GPUtraining target
Publish the evaluation (not the highlight reel)
Every Geocentric model ships with the evaluation that produced it: the cases it passes, the cases it fails, and the measurements behind both.
Capability is only half the result. A model we cannot characterise is a model we do not release.
