Video model

Photon 1 (Induction Labs)

Photon-1 is a sparse 106B-A5B mixture-of-experts imagination model trained on 18 years of computer use video. It predicts future states in a learned latent space via next-latent-token prediction, implicitly learning to act without action labels during pretraining. After reinforcement learning, it performs computer use tasks, simulates desktop and physical environments, and generalizes to games like checkers and billiards, outperforming a larger production language model on internal benchmarks.