OWLSimulate Anything on Anything
Open science world models that run on consumer hardware. From mobile phones to gaming PCs, experience state-of-the-art AI models at blazing speeds.
Running Live on Consumer Hardware
Our models achieve unprecedented performance on everyday devices. No cloud computing required.
Real-time World Generation
Robotics Deployment
Benchmarks updated weekly •Try our evaluation platform
Latest Research
Deep dives into our latest breakthroughs and findings
Why Human Evaluation is the Missing Piece in World Model Development
We're releasing OWL Eval, the first open-source evaluation platform for AI-generated videos. Human evaluation reveals critical model failures that automated metrics miss.
Optimizing Diffusion Decoders with Depth Pruning (Using ODE Regression)
Delete up to half the layers in diffusion decoders and regress to match pre-pruned outputs for nearly no loss in quality - achieving 3.5x speedups.
Generation vs. Reconstruction: Striking A Balance
GANs on VAEs with small latents spoil the latents. Train encoder beforehand then pretend your decoder is a generator. Temporal consistency breakthrough.