Research
Runway's GWM Worlds 2 Turns World Models Into Steerable, Real-Time 3D Environments
Runway's second-generation general world model generates continuous 720p video and synchronized audio in real time, letting users and AI agents steer environments through free-form text rather than scripted clips.
Mainstay Communications Team · · 5 min read
In brief
On September 3, 2026, Runway released GWM Worlds 2 as a research preview, the latest iteration of its general world model family and its most direct bid for the robotics and embodied-agent simulation market. The model generates interactive environments at 720p and 24 fps with 48,000 Hz audio, accepts persistent world definitions and a timestamped stream of text actions, and can hand control to autonomous agents, making it the most feature-complete real-time world model the company has released.
What GWM Worlds 2 Does
Runway announced GWM Worlds 2 on September 3, 2026, as a research preview available via contact form. The model generates interactive virtual environments in continuous 720p video at 24 fps, with synchronized audio running at 48,000 Hz that covers speech, sound effects, and ambient noise.
Unlike Runway's prior video models, GWM Worlds 2 has no preset session length. The world continues generating from each new input rather than producing a fixed clip, so a user can navigate, redirect, or restructure a scene indefinitely.
The WorldPrompt Input System
The model accepts a two-layer input format Runway calls WorldPrompt. The first layer is a persistent genesis prompt: a scene description, named subjects with defined attributes, physical laws (gravity, collision behavior, camera perspective), and an optional first frame for visual grounding.
The second layer is a timestamped action stream. Users and agents address specific subjects or the scene itself with overlapping text commands, alongside per-frame camera inputs covering translation and rotation. The system supports first-person and third-person perspectives, multiple movement types (walking, driving, riding, flying), object interactions, and spoken dialogue between named characters with generated lip-synced speech.
The Robotics and Agent Case
Runway's research page makes the embodied-agent use case explicit: "GWM Worlds 2 can be used as a simulated environment for evaluations of agents, and for training them in diverse environments." In one demonstration, the model generated a scene containing a humanoid robot and two flags, and an external agent issued navigation instructions that the robot character then carried out within the generated environment.
Agents can control any subject defined in the WorldPrompt, meaning the same model that serves human players in a real-time session can run headlessly under automated control for training or evaluation pipelines.
Multiplayer and Authoring Modes
GWM Worlds 2 supports multiplayer sessions through LiveKit, with each client assigned a role: player 1, player 2, or director. The director role can steer the environment's rules while the player roles control subjects, giving teams a structured way to co-author or stress-test a world in real time.
The model also supports three distinct generation modes. Ahead-of-time mode accepts a fully timestamped event stream and produces a complete scene, useful for filmmaking and advertising. Turn-based mode pauses generation at decision points, suited to visual novels and interactive film. Real-time mode generates continuously from live input with minimal latency.
Runway notes that LLMs can author WorldPrompts from simple natural-language descriptions, lowering the barrier for teams that don't want to hand-craft the full input schema.
Architecture
The model uses an autoregressive diffusion architecture conditioned on three context types: a global context window holding the genesis prompt and first frame, current frame inputs combining camera and text actions, and cached past frames stored in a sliding KV-cache. Video and audio decoders run causally with that cache, which Runway says is what makes real-time decoding tractable.
GWM Worlds 2 is built on top of Runway's foundational audio-visual generation model, itself introduced as part of the GWM-1 family in December 2025.
GWM-1 to GWM Worlds 2
Runway released GWM-1 on December 11, 2025, as a three-variant family: GWM Worlds for explorable environments, GWM Robotics for synthetic training data and policy evaluation, and GWM Avatars for conversational characters. The initial GWM Worlds variant generated real-time interactive scenes but didn't include generated audio or the explicit WorldPrompt control system.
GWM Worlds 2 adds native audio generation, the two-layer WorldPrompt input format, multiplayer, and agentic handoff, while running on the same autoregressive diffusion backbone.
Where It Sits in the Market
The world-model space in 2026 has split along a clear axis: models that generate real-time steerable video (Runway, Google DeepMind's Genie 3) versus models that output exportable 3D geometry (World Labs's Marble, which generates editable Gaussian-splat scenes). NVIDIA's Cosmos platform targets a third approach, producing photorealistic synthetic video for robotics and autonomous-vehicle training pipelines, and has seen roughly 2 million downloads according to NVIDIA figures.
Runway's control surface for GWM Worlds 2, with its composable action spaces and explicit subject addressing, differs from the single-text-prompt steering used by earlier systems including Genie. The tradeoff is acknowledged: the model trades some fidelity for speed, rapid camera rotations degrade texture and geometry, and long-term scene consistency remains imperfect. No image references beyond the initial first frame are supported in the current research preview.
Company Context
Runway closed a $315 million Series E round in February 2026 at a $5.3 billion valuation, led by General Atlantic with participation from Nvidia, Fidelity, AllianceBernstein, Adobe Ventures, and others. The company said at the time it would use the capital to "pre-train the next generation of world models and bring them to new products and industries," with gaming and robotics named as priority expansion sectors beyond the media and entertainment base it built on Gen-2 and Gen-3 video.
GWM Worlds 2 is available via a contact form on Runway's research site. No public pricing or self-serve access has been announced.
Sources
- Runway Research: Introducing GWM Worlds 2, Runway
- Runway's GWM Worlds 2 Streams Interactive AI Worlds at 720p in Real Time, AlphaSignal
- GWM Worlds 2: Runway's Real-Time Interactive World Model, ActTwo
- Runway Research: Introducing Runway GWM-1, Runway
- Runway's GWM-1 Models Generate Videos with Consistent Physics, DeepLearning.AI The Batch
- AI Video Startup Runway Raises $315M at $5.3B Valuation, TechCrunch
- Introducing GWM Worlds 2, a Playable World Model, Runway on YouTube
Companies mentioned
Mainstay Digital · Spatial and 3D visualization for industrial companies