Induction Labs says action labels are a major constraint for agents that learn from video. The company has released what it calls imagination models, a foundation-model architecture designed to pretrain on raw video without labels identifying the action that produced each frame.

Its test system, Photon-1, is a sparse 106B-A5B mixture-of-experts, or MoE, transformer trained on what Induction Labs describes as 18 years of computer demonstration video. On an internal computer-use benchmark, the company reports that Photon-1 outperforms Gemini 3.1 Flash-Lite while using substantially less pretraining compute and costing roughly 3× less to serve. The work targets a core problem for computer-use agents: large amounts of screen video exist, but most of it does not come with structured records of the clicks, keystrokes, and instructions that produced each state.

How the imagination model works

An imagination model predicts future frames autoregressively through a next-latent-token-prediction objective. During pretraining, it does not generate pixels directly. Instead, the system models video in a learned representation space.

The central claim from Induction Labs is that predicting future states can teach a model to complete tasks even when it never observes explicit action labels during pretraining. The company describes this learned behavior as an implicit policy. Rather than learning a label for every mouse click, the model is intended to learn higher-level concepts of what a person is doing on a computer.

Compression and representation

Photon-1’s architecture relies on a vision encoder using finite scalar quantization, or FSQ. Each frame is compressed into 960 discrete tokens. Each token is an 8-dimensional vector, and each dimension can take one of five values: −1, −1/2, 0, 1/2, or 1. That produces a codebook with 5⁸ possible codes.

The resulting representation is approximately 2.2 KB per frame. Induction Labs reports that this provides more than 100× better compression than existing OCR and multimodal-model representations while preserving text, layout, and state changes. That compression is central to the claim because video pretraining is constrained not only by model size but also by the number of frame tokens that can fit into a practical training budget and context window.

To reach that compression rate, Photon-1 uses a differential latent encoder. Instead of encoding each frame only as a standalone image, it encodes video frames as pairs, so the latents describe differences between frames rather than only frame contents.

Dataset and pretraining compute

The training corpus began with an internal index of 2 billion publicly available videos. Filtering reduced that collection to about 2 million computer screen recordings. Induction Labs then used an internal keyframe detection model to remove redundant frames.

The final dataset contained 575 million frames sampled at 1 frame per second. The company says this corresponds to 552 billion tokens, or about 18 years of video. Photon-1 was pretrained from scratch for a single epoch.

Training the 106B-A5B MoE model with a 32K context length took about 30,000 H200 GPU-hours, or 4.4×10²² training FLOPs. The research team implemented the training stack in PyTorch and used custom fused kernels for the vision encoder and MoE layers, sustaining 40% end-to-end MFU. The reported figures are consistent with each other: 30,000 H200-hours at 40% MFU is close to 4.3×10²² FLOPs.

Moving from prediction to action

After pretraining, Induction Labs finetuned Photon-1 on fewer than 35,000 computer-use trajectories to teach the action and instruction format. Special computer-use tokens allow the model to emit actions. During inference, Photon-1 first predicts the next frame’s state and then outputs the action needed to reach that state.

The company then applies online reinforcement learning. Rollouts run in real time on virtual machines at scale, and outcomes are programmatically verified to generate reward. The Linux virtual machines include five desktop environments: LXQt, Xfce, MATE, GNOME, and Plasma. Each VM has a Google account for web applications that require login access, as well as an internal ChatGPT clone with no rate limits. This setup is meant to test closed-loop behavior rather than only offline prediction, since a computer-use model must recover from its own actions and operate across changing interfaces.

Compute and cost comparison

Induction Labs says its compute and cost comparisons are weighted at a 10:1 input-to-output token ratio, which it says matches its computer-use tests.

Two caveats apply to those figures. First, the Gemini comparison is Induction Labs’ own conservative estimate, assuming 8B active parameters and 25T pretraining tokens. On those assumptions, the ratio is about 27× rather than the 30× headline figure. Induction Labs says “at least 30×” because it believes the real Gemini figure is likely higher and that the model was likely distilled. Second, the benchmark is internal and has not been released, so the result is not independently reproducible at this time.

Photon-1’s own breakeven cost on Induction Labs’ hardware is reported as $0.06 per 1 million input tokens and $0.60 per 1 million output tokens, without speculative decoding.

Tests beyond desktop video

Induction Labs also evaluated whether Photon-1 could generalize beyond desktop computer-use video, the only kind of video it saw during pretraining. The research team finetuned the model on domains absent from pretraining and compared it with two baselines: a vision encoder baseline with the same architecture and size but no imagination pretraining, and an LLM baseline, Ling-flash-2.0 from Inclusion AI, which was pretrained on 20T tokens.

On 20,000 tournament checkers games from the Open Checkers Archive 2.0, Photon-1 outperformed both baselines on world simulation and move quality. On 10,000 synthetically generated billiard games simulated at 5 fps, Photon-1 produced a mean absolute error of 0.47 against the ground-truth physics engine, compared with 1.15 for the LLM baseline and 1.44 for the vision encoder baseline. These tests do not replace public computer-use evaluation, but they are intended to probe whether the pretraining objective learned transferable state dynamics rather than only desktop-specific visual patterns.

Induction Labs also says Photon-1 acquired human priors from the pretraining video. After reinforcement learning, the model learned to use the in-VM ChatGPT clone to draft artifacts and answer knowledge questions, steering the LLM in a way comparable to a human user.

The company has not released model weights, an API, or a license for Photon-1. The result remains a research demonstration rather than a deployable public model. Induction Labs also published an announcement thread on X.