Vertex AGI exists on a simple bet: most tasks don't need a frontier model, they need a model that's good at exactly one thing, cheap enough to run locally, and honest about where it falls short.

That means every release ships with a held-out evaluation — never a benchmark sampled from the same pool used to train it — and a limitations section that says plainly what the model gets wrong, not just what it gets right. We'd rather publish a model that trails a larger competitor by six points, with the numbers to prove it, than round up.

It also means building on hardware you could actually buy. Every fine-tune in the Prism, Amethyst, and Copal families runs on Apple Silicon. The from-scratch mixture-of-experts pretrain described in our research is training on a single 6GB gaming GPU in a Windows desktop — not a cluster, not cloud credits.

We publish weights, not just claims. Every model we ship is on Hugging Face in both MLX and GGUF formats, so anyone can verify what we say about it.