Reka releases Apache-2.0 inverse-dynamics weights for interactive world models

Reka AI has released the weights and code for its Inverse Dynamics Model (RIDM), a pair of small video models designed to infer camera actions from observed motion. The release is intended as a component for interactive world models: instead of predicting the next image directly, RIDM works backward from visual change and estimates actions such as W/A/S/D/Shift inputs together with yaw and pitch. Reka publishes separate optical-flow and raw-pixel variants, allowing builders to trade a heavier preprocessing dependency for different generalization behavior.

The artifact layer is unusually inspectable. The official Hugging Face repository includes `model.safetensors` checkpoints for both variants, JSON configs, inference scripts, smoke tests, expected sample output and an explicit Apache-2.0 license for the weights and code. Reka lists 1,795,337 trained parameters for the flow model, plus a frozen 990,162-parameter RAFT-small component, and 9,836,063 parameters for the pixel model. The external RAFT-small weights remain under BSD-3 and are not bundled as part of the Apache-licensed model artifacts.

Reka says both variants were trained from 4,071 recorded gameplay sessions. The flow training set contains 7,071,037 frame pairs, while the pixel model used roughly 104,000 gameplay windows; both were trained on a single NVIDIA L4 GPU. Evaluation percentages in the repository are publisher-reported and remain attributed as such. For Open Model Weights, the release is relevant as a reproducible open-weight robotics/world-model component, but the flow and pixel checkpoints should remain distinguishable rather than being collapsed into one generic model record.