Announcing Geocentric Kestrel.

Geocentric Kestrel is our flagship model and the largest we have trained. It has 249,396,480 parameters, about 250 million, and it was trained on 5,720,702,976 tokens on a single RTX 2060.

Kestrel is built around the part of computing that stays put: the Linux kernel, the C that runs it, and the networks it talks over. It learned from 57 programming languages, the entire Linux kernel, more than 80,000 man pages and a great deal of C, alongside networking and cybersecurity material and general text from Cosmopedia, FineMath, FineWeb and FineScience. Linux, C and networking are where it is strongest, and it can detect flaws in code, configurations and systems.

Why you cannot download it

We are not releasing Kestrel's weights. A model that can find flaws can help the people fixing them and the people exploiting them, and open weights cannot be recalled once they are out. Because of those security and safety concerns, Kestrel is announced but not publicly available, and there is no hosted version.

Arc, our compact co-writing model, remains open. Its weights are on Hugging Face and Ollama under the Apache License 2.0, and its complete evaluation, including every response it gets wrong, is on its model card.

What we have not published yet

We have not published evaluation results for Kestrel, and we are not making claims about how it compares with other models. Its context window is 1,024 tokens, and a later Kestrel release will have a much longer one.