For a long time, I’ve had the feeling that generative AI models, for most of the tasks they’re used for, don’t actually need the oversized capabilities offered by frontier models. On top of that, they’re often prohibitively expensive when accessed via API. An open-source model with 4–14 billion parameters, multimodal capabilities, and the right harness of tools, data, and knowledge is more than enough for most use cases. With this idea in mind, I decided to build, together with my team, a si...