Runway announces GWM-1, its first world model

Background

A world model is a system trained to predict how an environment changes over time: given a current state of a scene and something that acts on it, the model estimates what comes next. This differs from the way language models work, since the object being predicted is the state of a simulated world rather than the next word in a sentence. Runway said it had introduced the General World Models research direction two years before GWM-1 ^[2]. The company argued that language models alone would not solve robotics, disease or scientific discovery, and that world models offered the clearest path to general-purpose simulation ^[2]. That argument frames the announcement less as a single product launch than as a claim about which kind of model the next stage of AI research will rest on.

What was announced

On 11 December 2025, Runway announced GWM-1, described as a general world model that simulates environments in real time ^[1]. The model comes in three variants: Worlds, Avatars and Robotics ^[1]. It is built on Gen-4.5, and it generates 720p video for clips of up to 2 minutes ^[1]. Alongside GWM-1, the company announced native audio and multi-shot editing for Gen-4.5 ^[1].

Key details

The three variant names — Worlds, Avatars and Robotics — point to three different settings in which a simulated environment is useful, and the announcement presents them as one family built on a shared model rather than as separate products ^[1]. The technical figures given are concrete: 720p resolution, and a maximum clip length of 2 minutes ^[1]. Resolution and clip length matter for simulation because they bound how much detail a scene can hold and how long a single consistent episode can run, which in turn limits what a user or a robot policy can learn from it.

Real-time operation is the other stated property ^[1]. Generating frames quickly enough to keep pace with events as they happen is a different requirement from producing a polished video after the fact: it means the simulated environment can respond while someone is interacting with it, which is the property that matters if a model is to be used as a stand-in for the physical world during training or testing.

GWM-1 sits on top of Gen-4.5, the company's existing generation model, and the same announcement extended that base model with native audio and multi-shot editing ^[1]. Native audio means sound is produced as part of generation rather than added afterwards, and multi-shot editing means a sequence can be assembled from more than one shot within the same workflow ^[1].

Why it matters

World models are attractive because physical and biological systems are expensive or impossible to experiment on directly. If a model can simulate an environment faithfully and quickly, it can serve as a place to try things out: a robot can be trained or evaluated without repeatedly risking real hardware, and a scientific or engineering idea can be explored before committing resources to it. Runway's stated position is exactly this — that language models alone would not solve robotics, disease or scientific discovery, and that world models offer the clearest path to general-purpose simulation ^[2].

The two-year gap the company draws between introducing the research direction and shipping GWM-1 places the release as an attempt to turn a research agenda into a working model family ^[2].

World models as a class still face open problems: keeping a simulated scene consistent over long stretches, getting physical behaviour right rather than merely plausible, and covering the rare situations where systems actually fail. The announcement should be read with those difficulties in mind, and with attention to how the model behaves once people outside the company use it.