Evoke open-sources a 14B real-time interactive world model

Evoke is an open-source AI world model that turns a single still image into a playable, steerable interactive environment in real time. Feed it an image, plug in a joystick, and you can pilot a snowmobile, kayak through rapids, or go scuba diving while the model streams video back to you on the fly.

What you can do with it

Evoke turns static visuals into continuous, interactive simulations. Your joystick controls movement and camera rotation, and the video immediately updates to match your path.

Crucially, you can re-prompt the scene mid-flight without cutting or restarting the session. If you are cruising down a road and type "add balloons to the sky" or "trigger an aurora borealis," the model smoothly weaves the new elements directly into the live rollout.

Beyond toy demos and game-like exploration, its primary utility lies in robotics. Because it understands diverse physics and extreme environments, developers can use it to generate synthetic training runs for robots in simulated hospitals, underwater expeditions, or disaster relief zones.

How it works and hardware requirements

The secret to its speed is distillation: Evoke generates video in just three steps with zero classifier-free guidance (CFG), meaning it requires only one forward pass per step. On a single Nvidia H200 GPU, it generates 1.5 seconds of 384×640 video every 2.11 seconds. Clustered across five H200s, latency drops to 1.34 seconds.

Unlike traditional diffusion models that crash into context limits after a few dozen frames, Evoke is designed to run continuous sessions for hours. It stores scene geometry in an external, camera-indexed point cloud memory bank rather than cramming previous frames into the denoiser context. As a result, memory consumption stays bounded regardless of rollout length.

On the interactive world benchmark WBench, Evoke secured state-of-the-art marks while remaining competitive on visual quality benchmarks like VBench-Long.

The model has 14 billion parameters, and the final weights weigh in at roughly 57GB. AlayaLab has published the code, UI launchers, and intermediate training checkpoints under the Apache 2.0 license on GitHub (AlayaLab/Evoke). Setting it up requires Python 3.10, CUDA 12.4, DeepSpeed, and Flash-Attention.

Why it matters

While frontier labs have teased closed systems like Genie 3, Evoke is the first genuinely capable, long-horizon interactive world model that anyone can actually download and audit down to its raw weights.

The catch is hardware: a 57GB footprint means home enthusiasts on consumer GPUs won't be running this locally anytime soon. But for robotics labs and simulation teams with access to enterprise silicon, this is an immediate, high-leverage sandbox worth spinning up right away.