CADRE 2x16x16 world models
This release contains four separately trained video world models. Each predicts future video from context frames and a sequence of actions, using the original CADRE 2x16x16 autoencoder.
Research release
Each environment defines a different action space
Robot manipulation uses 14 joint-state deltas for two YAM arms. Driving uses speed and curvature. Indoor navigation uses six camera-motion values. Human manipulation uses 102 hand-pose deltas. These inputs are not interchangeable.cadre-world-models-2x16x16: Kyutai model card
One latent step predicts two frames
At the documented 20 frames per second, each predicted latent step corresponds to 100 milliseconds of video. The included autoencoder is the existing original checkpoint, not a fifth world model.cadre-world-models-2x16x16: Kyutai model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | kyutai/cadre-world-models-2x16x16robot_manipulationdrivingindoor_navigationhuman_manipulation |
| Model architecture | 1.2B-parameter causal diffusion transformer per environment.cadre-world-models-2x16x16: Kyutai model card |
| Terms | CC BY-NC 4.0; the included autoencoder also uses DINOv3 components.cadre-world-models-2x16x16: Kyutai model card |
Related models
- CADRE 2x16x16
Kyutai · CADRE
A causal video autoencoder with two-frame temporal compression and a 16-by-16-pixel spatial grid.
Launch date not confirmedResearch releaseDownloadable - CADRE action probes
Kyutai · CADRE
Four inverse-dynamics models that estimate actions from generated frames to evaluate CADRE world models.
Launch date not confirmedResearch releaseDownloadable