Runway, the AI company behind cinematic video-generation models such as Gen-4.5, expanded into world models with the launch of its first simulation-focused system in December and plans for another release later this year. The New York–based startup is training models on observational video data to better simulate how real-world environments behave, positioning video-based intelligence as a path beyond text-centric AI systems.
The company has grown into a global operation with offices across New York, London, San Francisco, Seattle, Tel Aviv and Tokyo and commercial partnerships with media companies including Lionsgate and AMC Networks. Runway also raised $315 million in February and works with compute partners such as Nvidia and CoreWeave while applying its multimodal AI systems to robotics and simulation research.
For creators and researchers, Runway’s world-model strategy could enable faster experimentation across filmmaking, robotics and scientific workflows by using AI-generated simulations instead of physical testing environments. The approach reflects a broader industry shift toward multimodal AI trained on real-world sensory data rather than language alone.
Video-Centric World Models
Runway Launches Its First World Model
Trend Themes
-
Video-centric World Models — By training on observational video data, models can recreate dynamic real-world environments, enabling synthetic yet physically coherent scenarios that challenge text-only AI assumptions.
-
Multimodal Sensory AI — Combining visual, temporal and contextual cues offers richer representations of environments, opening pathways to systems that reason across sight, motion and context rather than language alone.
-
Simulation-first Content Production — Generating high-fidelity simulated footage for testing and previsualization creates alternatives to costly physical shoots and prototypes, shifting value toward virtual experimentation.
Industry Implications
-
Film and Media — AI-driven world models can produce realistic scene simulations and virtual actors that redefine production pipelines and reduce dependency on physical locations and crews.
-
Robotics and Automation — Simulated environments grounded in real video data enable more transferable training for robots, decreasing reliance on time-consuming real-world trials and accelerating deployment cycles.
-
Scientific Research and Simulation — Data-rich video simulations can replicate complex physical systems for hypothesis testing, offering lower-cost, higher-throughput alternatives to many laboratory experiments.