Induction Labs has released Photon-1, a sparse 106B-A5B mixture-of-experts transformer that learns to simulate desktop environments and play games by predicting future frames from raw video without action labels. The model uses finite scalar quantization to compress each frame into 960 discrete tokens, achieving over 100 times better compression than existing representations while preserving text and layout details.
- Photon-1 was pretrained on 575 million frames (approximately 18 years of video) from a filtered corpus of 2 billion public videos.
- Training required approximately 30,000 H200 GPU-hours and 4.4×10²² FLOPs, using a next-latent-token-prediction objective.
- On an internal computer use benchmark, Photon-1 outperforms Gemini 3.1 Flash-Lite while using roughly 27 times less pretraining compute.
- After finetuning on fewer than 35,000 trajectories, the model beats baselines in checkers and models billiard physics with a mean absolute error of 0.47.
The research demonstrates that predicting future states teaches an implicit policy for task completion, though Induction Labs notes this is a research result with no weights or API available.