Three weeks after it quietly posted a research preview, Runway is explaining how it made a video model you can play. In a Latent Space deep-dive published Thursday, September 25, the company's CTO, its principal research scientist and its co-CEO described the engineering behind GWM Worlds 2, a world model that streams continuous 720p video at 24 frames per second with 48kHz audio for as long as the session runs. They also explained WorldPrompt, the input format Runway hopes will become the control layer for AI-generated worlds.
GWM Worlds 2 was announced on September 3, 2026. It follows the original GWM Worlds, which Runway showed in December 2025 as part of its first "General World Model" family built on Gen-4.5. The new version adds generated sound, including speech, sound effects and ambience, and much finer control over subjects and scenes. The world continues from each new input rather than playing back a fixed clip, so sessions have no preset length.
How WorldPrompt Works
WorldPrompt splits a world into what stays the same and what changes. The persistent layer is a "genesis prompt" that describes the environment, the subjects who can act, with their appearance and voices, and the "laws" that govern behavior, such as gravity, collisions and camera perspective. A first frame anchors it visually. On top of that sits a timestamped event stream: free-form text actions with start and end times, each addressed to a named subject or to the scene itself, plus a per-frame stream of camera movement. Actions can overlap, and speech is treated as one more action that carries the line to be spoken.
Runway describes three ways to drive the same model. Ahead of time, a user (possibly with an LLM's help) writes the whole event stream up front, which suits film and advertising. Turn-based, generation pauses at decision points for visual novels and interactive film. Real-time, actions affect the stream immediately, which suits games. Runway calls the real-time mode the hardest, because reactions have to land within tens of milliseconds, faster than anyone can type or than a vision-language model can respond. Its demo binds keys and mouse clicks to premade prompts instead.
Robin Kahlow, Runway's principal research scientist, told Latent Space the format is meant to "control all the different subjects in the world," much like a game engine. "Like, if there's an NPC somewhere, the NPC might walk up to you and say something," he said. "So you could achieve the same thing with this kind of model, where you can have very detailed control over everything in the scene." CTO Kamil Sindi put it more bluntly: "You can create promptable worlds on-demand with video and audio in sync, across all these different domains and environments. That's not a distant-future hypothetical thing."
From Clip Generator to Runtime
Under the hood, Runway takes its foundation audio-video model, fine-tunes it on WorldPrompt, then post-trains the bidirectional model, which generates a whole clip at once, into an autoregressive one that produces a frame or a few frames at a time. Distillation then makes it fast enough to run live. Co-CEO Anastasis Germanidis said on the podcast that step distillation can cut a model from about 50 denoising steps to four with "comparable outputs." Each step attends to the genesis prompt, the current inputs and a sliding cache of recent frames.
That design has costs. "The biggest challenge with autoregressive models is error accumulation," Germanidis said. "You're feeding generated frames back into the model to generate the next frames, and if there are any small errors, they accumulate over time." Sindi said deciding "what context to keep, what to discard" is a constant tradeoff "so we're not blowing up our GPU memory." Kahlow was frank about the gaps: "The model does not have perfect memory. That's still an open research problem." Runway's own post warns that fast camera turns can degrade textures and geometry. It also notes that the model has no structured world state, so agents and harnesses have to infer what happened from pixels and audio.
Why It Matters
World models are becoming the next contested frontier after chatbots, and headline specs are converging. Google DeepMind's Genie 3 also runs at 720p and 24fps, but Google says it currently supports "a few minutes of continuous interaction, rather than extended hours." World Labs, Odyssey, Decart and NVIDIA's Cosmos are all working on nearby problems. Runway's pitch is less about raw fidelity and more about control. WorldPrompt turns a single blob of text into named subjects, explicit laws and overlapping timed events, and it supports multiplayer roles such as player 1, player 2 and director, broadcast over LiveKit. That kind of structure is what a game studio, a robotics lab or an agent-evaluation team would need to build on.
There is also a business case. Runway, reportedly valued at $5.3 billion after a $315 million raise in February, has already found paying customers in robotics labs that use its GWM models for synthetic data and closed-loop policy testing. Germanidis argues the economics favor real-time generation too: "I'll be very surprised if we're, two years from now, we don't primarily use real-time models."
What to Watch
GWM Worlds 2 is still a research preview, with access only through a business contact form and no pricing. The next marker comes soon: Runway hosts its AI Summit in San Francisco on September 30, with speakers from DeepMind, NVIDIA and Physical Intelligence, and the event is focused mainly on physical AI. Watch for a timeline to broader access, for any move to publish WorldPrompt as an open spec that other tools can target, and for signs of the unified model Germanidis described, one that handles navigation, manipulation and human behavior in a single real-time system. Google's next Genie will show whether session length or control becomes the benchmark that matters.
“You can create promptable worlds on-demand with video and audio in sync, across all these different domains and environments. That's not a distant-future hypothetical thing.”— Kamil Sindi, CTO, Runway