Most video models focus on visual fidelity; this release focuses on aligning video synthesis with physical-world reasoning for embodied tasks. That shift matters because downstream robotics, simulation, and multi-step interaction tasks require temporally coherent videos that reflect plausible object dynamics and task completion, not just photorealism.
Key Capabilities
- MoE scale: a 30B-parameter Mixture-of-Experts backbone plus a refiner, designed to increase capacity while keeping inference throughput tractable for large-video generation scenarios. This design targets higher long-horizon and multi-entity reasoning compared with dense counterparts.
- Large embodied training signal: trained on a mixture that the authors report as including 70,000+ hours of web-sourced embodied video data, with multi-reward supervision for aesthetics, physical rationality, and task completion — meaning outputs are optimized for plausible interactions, not only appearance.
- End-to-end inference workflows: provides a structured prompt rewriter (two-stage rewriter + LoRA adapter), auto-negative pruning, and ready scripts for single- and multi-GPU inference across diffusers and SGLang backends, plus FSDP/CP8 sharding options for MoE checkpoints.
- Benchmarked leadership: reported top ranking on a public RBench leaderboard (as of early July 2026) on metrics combining manipulation, long-horizon reasoning, and embodied scenarios.
Who it's for — and tradeoffs
Great fit if you need research-grade video models that emphasize physical plausibility and task-oriented video generation (robotics simulators, embodied AI research, multimodal agents). The project is released under Apache-2.0 and bundles dense and MoE checkpoints plus rewriter components for production-style inference.
Look elsewhere if you need a tiny, single-GPU consumer model for quick mobile prototyping: MoE inference requires substantial system RAM and multi-GPU or specialized runtimes (SGLang/CP8/FSDP) for practical throughput. Expect engineering effort to set up the two-stage rewriter, LoRA adapters, and grouped-expert runtime for best performance.
