Toward V4: Our Own Foundation Models

We are working on our own GPT-style foundation models, pretrained from scratch, for Orion V4. V4 is coming soon, and this research is the next step in building more of Orion’s intelligence ourselves.

Building from the foundation

Pretraining is the stage in which a language model learns general patterns from data. Training from scratch means starting with newly initialized model weights and learning that foundation ourselves. It gives us a deeper role in shaping the model that will eventually power the product.

Our ambition is to bring the model and the app closer together: intelligence shaped around the conversations, writing, reasoning, and personal experiences people come to Orion for.

V3 today. V4 in development.

Orion V3 uses a Qwen 3.5-based model. Our own pretrained models are still in development for V4. This is a research preview; model sizes, evaluation results, and release details will come as the work takes shape.

For us, progress means models that are useful in practice: clear responses, dependable behavior, and an experience that fits the device in your hand. Those are the questions that will guide how we evaluate the work.

A personal place in a much bigger future

The prospect of artificial general intelligence and superintelligence makes this an extraordinary time to build. We want Orion to contribute to a future where people can use increasingly capable intelligence directly, personally, and privately.

V4 is the next chapter of that ambition. We are building, learning, and moving toward it.