Induction Labs Unveils Photon-1: An 'Imagination Model' That Learns Without Action Labels, Outperforms Gemini

Induction Labs is challenging a fundamental assumption in reinforcement learning and agent development: the necessity of explicit action labels paired with video frames. Traditionally, agents learn by observing a sequence of states and the corresponding actions that led to those states. Induction Labs argues this requirement acts as a bottleneck, limiting scalability and efficiency. To address this, they have introduced "imagination models," a foundation model architecture that undergoes pretraining solely on raw video, entirely devoid of action annotations.

The company's demonstrator for this novel approach is Photon-1, a sparse 106 billion parameter mixture-of-experts (MoE) transformer. This sophisticated model was trained on an extensive dataset comprising 18 years of computer demonstration videos. The breadth of this training data has enabled Photon-1 to acquire a remarkable range of capabilities, including the ability to simulate desktop environments, engage in strategic gameplay like checkers, and accurately model the physics of billiard balls. These diverse skills are achieved from a single, unified pretraining run.

In internal evaluations conducted on a computer use benchmark, Induction Labs reports that Photon-1 demonstrates superior performance compared to Google's Gemini 3.1 Flash-Lite. Crucially, Photon-1 achieves this by consuming significantly less pretraining compute and incurring approximately one-third of the serving costs associated with Gemini 3.1 Flash-Lite. This efficiency is a key differentiator, suggesting a more sustainable path for developing advanced AI agents.

The core mechanism behind the "imagination model" is its ability to predict future video frames autoregressively. This prediction is driven by a "next-latent-token-prediction" objective. Importantly, during the pretraining phase, the model does not generate actual pixels. Instead, all its learning and modeling occur within a learned, compressed representation space. The central claim advanced by Induction Labs is that this predictive capability, even in the absence of explicit action supervision, inherently teaches the model to complete tasks. They refer to this as an "implicit policy." The model learns to grasp the underlying concepts of human actions – such as the intent behind a mouse click – rather than simply memorizing discrete labels for each action.

To achieve the remarkable scalability and efficiency of this approach, Photon-1 employs a specialized vision encoder that utilizes finite scalar quantization (FSQ). This technique compresses each video frame into 960 discrete tokens. Each token is an 8-dimensional vector, where each dimension can assume one of five specific values: -1, -1/2, 0, 1/2, or 1. This configuration provides a codebook of 5⁸ possible unique codes per token. The resulting compressed representation is remarkably compact, measuring approximately 2.2 KB per frame. Induction Labs reports that this method achieves over 100 times better compression than existing representation techniques used in optical character recognition (OCR) and other multimodal models, while simultaneously preserving critical information such as text, layout, and state changes within the video. To further enhance this compression rate, Photon-1 incorporates a differential latent encoder. This encoder processes video frames in pairs, enabling the latent representations to capture the differences between consecutive frames rather than the entire content of each individual frame. This differential approach is particularly effective for modeling sequential data like video, as it focuses on the dynamic changes over time.