World-action models predict both what a robot should do next and what the scene will look like once it has done it. The strongest versions are built on video diffusion transformers, which are too heavy to run on hardware mounted on the robot itself. A group publishing as ML Collective has released SmolWAM, a 0.88-billion-parameter model derived from NVIDIA’s two-billion-parameter Cosmos Policy, keeping the pretrained trunk but rebuilding the inference path around it.
The changes are mostly about doing less work at inference time: single-frame visual latents, cached text features, using 12 of the model’s 28 transformer blocks, and reading out both actions and future visual latents in a single rectified-flow step.
The reported figures come from the authors’ own evaluation and have not been independently reproduced. They are 95.5% success on the LIBERO benchmark across 400 episodes, against 98.2% for the teacher model; 57 milliseconds per 16-action chunk; 2.4 GiB of GPU memory; and 3.6 GPU-hours of training. Against Cosmos Policy they claim 11 times faster inference and 18.5 times fewer floating-point operations.
This is a NeurIPS 2026 workshop paper, so it has had light review rather than none. Code is on GitHub under Apache 2.0, though the weights are derived from NVIDIA’s Cosmos Policy checkpoint and inherit its non-commercial licence, which rules out commercial use for now.