Runway has released GWM Worlds 2, a research preview that transforms video and audio generation into real-time interactive simulations you can control by typing prompts. The tool uses a new input format called WorldPrompt, which lets you set an initial scene and then add timestamped actions that play out in real time, rather than generating entire video clips at once.

What Makes Runway's Approach Different From Competitors?

Runway, valued at $5.3 billion after raising $315 million in February 2026, is positioning GWM Worlds 2 as a differentiator in a crowded field. The company faces direct competition from Google DeepMind's Genie 3, Odyssey-2 Pro, and World Labs' RTFM, all of which are pursuing similar goals of turning generative models into interactive simulation engines. What sets Runway apart is its focus on making generation fast enough to play in real time, rather than as a pre-rendered video.

The engineering challenges are substantial. The Runway team had to solve two critical problems: making the model generate frame by frame instead of producing an entire clip upfront, and ensuring the generation speed matches real-time playback. To achieve this, Runway fine-tuned its foundational audio-video model, added post-training for autoregressive generation (a technique where each new frame depends on the previous one), and used distillation methods to compress the model's complexity. According to the team, a model might go from around 50 denoising steps, which are computational passes that refine the output, down to just four steps, with some quality loss but potentially comparable results.

How Could Developers Actually Use This Technology?

The practical applications extend well beyond gaming. While interactive games are the obvious use case, the technology opens doors for robotics engineers and AI researchers testing agents at scale. Instead of building thousands of unique simulated environments by hand, developers can now generate them on demand by describing what they want. This dramatically reduces the time and resources needed to test how robots or AI agents behave in different scenarios.

Ways to Apply World Models in Development and Testing

• Game Development: Developers can create interactive environments on demand without manually building each scene, allowing rapid prototyping and iteration of game mechanics and worlds.
• Robotics Testing: Engineers can simulate how robots respond to different environments and scenarios, testing behaviors before deploying physical hardware in the real world.
• Agent Training at Scale: AI researchers can generate thousands of diverse simulated environments to train and test autonomous agents, which is far less resource-intensive than creating each environment manually.
• Synthetic Data Generation: The tool can produce training data for machine learning models, allowing researchers to generate labeled examples of specific scenarios without collecting real-world footage.

Runway's CTO Kamil Sindi noted that scaling data and models is resulting in better performance, but the research preview still has significant limitations. The company acknowledges several open research problems that must be solved before GWM Worlds 2 becomes a production-ready tool.

What Are the Current Limitations Holding Back Deployment?

Despite the promise, GWM Worlds 2 remains a research preview with acknowledged flaws. Reliability depends heavily on how difficult the requested action is; simple movements work quite reliably, but more complex scenarios may fail. The model also lacks perfect memory, meaning it can lose track of details or context over long sequences of interactions, which Runway describes as an open research problem.

Error accumulation poses another challenge. In autoregressive models, where each frame depends on the previous one, small mistakes can compound over time, degrading quality as the simulation runs longer. Managing GPU memory during infinite generations is also non-trivial, as the model must maintain state without running out of computational resources. Perhaps most intriguingly, the team identified a problem called counterfactual generation: training data often contains biases. For example, football training data contains far more successful goals than failed attempts, so the model might render success more convincingly than failure. Until these gaps close, the fully self-generated, real-time games that world models promise will likely remain a research goal rather than a shipping product.

"It's a research preview and not perfect," noted Kamil Sindi, adding that reliability depends on how difficult the action is; movement works quite reliably.

Kamil Sindi, CTO at Runway

Runway's journey to GWM Worlds 2 reveals how the company evolved from its roots. The company began as a platform to make open-source models usable for artists, then spent its early years on a segmentation tool called Green Screen that was famously used in the film "Everything Everywhere All at Once." The pivot to generative video came in 2022, when Runway signed a deal for a cluster of 1,000 A100 graphics processing units (GPUs), which are specialized chips designed for AI training, and released Gen-1 in January 2023. This infrastructure investment laid the groundwork for the more ambitious world model research that followed.

The race to build practical world models is heating up across the AI industry. Runway's $5.3 billion valuation reflects investor confidence that generative simulation will become a major category, but the company must overcome significant technical hurdles before GWM Worlds 2 moves beyond research status. The next phase will likely depend on whether Runway can solve the memory, error accumulation, and bias problems that currently limit the tool's reliability and scope.