Sber open-sources Kandinsky WM 1.0 world models
Sber has released Kandinsky WM 1.0 under an MIT license: three video-generation world models for driving, robotics and physics-heavy scenes, built to produce physically plausible synthetic training data for Physical AI.
Sber's Kandinsky team announced Kandinsky WM 1.0, a set of three specialised video-generation models: one for automotive scenarios, one for robotics and one for scenes with complex physics. Code and weights are available under the MIT license on GitHub, GitVerse and Hugging Face.
The release follows Kandinsky 5.0, whose flagship Video Pro model the team says held first place among open-source text-to-video models on arena.ai for an extended period.
The team is explicit about what WM 1.0 is and isn't. Using the World Labs taxonomy, which splits world models into renderers, simulators and planners, it classifies WM 1.0 as a renderer. Each model takes a first frame and a text prompt and generates a visual continuation of the scene. It does not predict world state or choose actions for an agent.
The stated goal is synthetic data for Physical AI, where real-world collection is slow and the rare, dangerous cases matter most. The authors note that a single car recording around the clock for a year captures fewer than 10,000 hours of road footage. By comparison, NVIDIA trained Cosmos on 20 million hours of video and Meta trained V-JEPA 2 on more than 1 million hours. They argue that synthetic data only helps if the generator preserves geometry and physical dynamics. Otherwise it adds model errors to the dataset.
According to the release materials, the models have about 2 billion parameters each. They place in the top 10 on PAI-Bench-G, Physics-IQ and RBench and edge out NVIDIA's Cosmos-Predict2.5-2B in their target domains, but they still trail much larger systems.