Photon-1
#6 en Montage vidéo IAInduction Labs · v1 · 3× · vu le 28 juil. 2026
Photon-1 is not a video-editing tool, but a research Foundation Model from Induction Labs, referred to as an "Imagination Model." It learns by predicting future video frames in a latent representation space from unlabeled screen recordings, without ever seeing action labels. Photon-1 is a sparse 106-billion-parameter Mixture-of-Experts Transformer model with 5 billion active parameters, trained on approximately 575 million frames (roughly 18 years) of computer screen recordings. According to Induction Labs, it outperforms the Gemini 3.1 Flash-Lite model on internal computer-use Benchmarks with significantly lower training effort. It is explicitly a research result with no published weights, API, or license.