Open-source visual reasoning model matches proprietary AI on diverse tasks without locked-up training data

Researchers released Vero, a family of open visual reasoning models that perform as well as or better than closed proprietary competitors across 30 different benchmarks. They built this by scaling reinforcement learning training across 600,000 examples from 59 different datasets and releasing all code, data, and model weights publicly.

For years, the strongest visual reasoning models have been locked behind proprietary systems with closed training recipes and private datasets — which means researchers outside those companies couldn't understand how they work or build on them. This paper shows that the techniques aren't secret; they're just expensive and tedious. Open-source models can now match proprietary ones, which means the knowledge about how to build visual reasoning systems is no longer gatekept by a handful of labs. What changes: any researcher or company can now replicate the strongest visual reasoning capabilities using public code and data.