Induction Labs has introduced Photon-1, a 106-billion parameter mixture-of-experts model trained on raw video without action labels. This 'imagination model' architecture predicts future frames in a learned representation space, enabling it to infer actions implicitly. Photon-1 reportedly outperforms Gemini 3.1 Flash-Lite on an internal benchmark, using significantly less pretraining compute and incurring lower serving costs. The model achieves high compression rates by encoding frame differences, and was trained using PyTorch on 575 million frames, demonstrating efficient use of H200 GPUs. AI
IMPACT Sets a new direction for agent training by demonstrating task completion without explicit action labels, potentially reducing data requirements.
RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
