Alaya Lab has released Evoke, a 14-billion-parameter open-weights world model that moves a generated scene's memory out of the image generator and into a separate store indexed by camera position. That design change is what lets the model keep a world consistent across hour-long interactive sessions instead of a few minutes, and the lab published not just the finished model but every checkpoint along its training path.
Key facts
- Evoke generates 1.5 seconds of video in 2.11 seconds of compute on a single H200 GPU at 384 by 640 resolution -- close to interactive, but still slower than playback.
- The release landed on August 23, 2026 and immediately took the number one slot on Hugging Face's daily papers list.
- Built by Alaya Lab; code and weights are Apache 2.0, with non-commercial exceptions for the depth backends.
- Primary source: the Evoke project page, with the paper at arXiv 2608.13546.
Every interactive video model so far has had the same embarrassing failure. Turn the camera away from a chair, turn back, and the chair is a different chair. Google DeepMind's Genie 3 -- the best-known system of this kind -- runs at 20 to 24 frames per second at 720p but supports only a few minutes of continuous interaction, because everything the model knows about the scene has to live inside the small window of recent frames it can attend to. Push past that window and the world quietly forgets itself.
Evoke's answer is to stop asking the generator to remember. In the project's own words, it "decouples world state from generation: persistent state lives beyond the denoiser and is addressed through camera pose." In practice, the model generates a chunk of nine latent frames -- about 1.5 seconds of video -- then estimates the depth of what it just drew, converts those pixels into three-dimensional geometry, and files that geometry away in an external bank. When the camera later points back at a place it has been, the bank is queried by camera position, ranks its stored views by how much they overlap with the current shot, returns up to eight of them, and hands the generator a warped picture of what should be there plus a mask showing which parts are genuinely unknown.
The analogy is a film crew that keeps a continuity binder. The camera operator does not have to remember what was on the desk in scene four; someone else wrote it down, and hands the page back when the scene is revisited. The generator's job shrinks to rendering, and rendering is something a very small number of denoising steps can do.
That last part is the second half of the trick. Evoke ships as a three-step model with no classifier-free guidance, which is roughly an order of magnitude cheaper per frame than a standard video diffusion model. Getting a three-step student to behave well over a long interactive rollout normally fails, because the student is trained on short clips and then asked to run for an hour. Alaya Lab rebuilt the teacher to make long-horizon supervision affordable: chunked sparse attention for local context, retrieval of distant frames, and a linear-attention global state, so the cost of scoring a long rollout grows roughly in proportion to its length instead of with its square. The teacher then grades a full twenty-chunk rollout -- about 31 seconds of continuous generated video -- rather than a short sliding window. If you want the background on why turning a slow, careful model into a fast one is hard, our lesson on diffusion distillation covers the mechanics.
The unusual generosity of the release is that the training ladder itself is public. The Hugging Face repository exposes the teacher bundle and separate checkpoints for camera control, few-step training, long distillation, and post-distillation. Most labs publish the endpoint; publishing the rungs lets other researchers pick up the method partway rather than reproducing it from scratch. The GitHub repository carries the Apache 2.0 license, with the README carving out vendored third-party code, and the model card warning that the depth backends must be downloaded separately under a non-commercial license.
You can also talk to the world while it runs. Evoke accepts prompt changes mid-session, and the paper is honest about the asymmetry: adding something new that has not yet been pinned to geometry usually works, while overwriting scenery the bank has already anchored usually does not. Text can introduce events into the world; it cannot easily repaint history.
Why it matters: an interactive world that stays consistent for an hour is the missing ingredient for training embodied agents in simulation, and this is the first open-weights system in that class -- a counterpart to closed products like Vidu's streaming service. The wider trend is visible elsewhere too; three papers in a single day recently argued that the training world matters more than the model.
The honest caveat is the memory ceiling. Evoke's long-session setup retains roughly 90 seconds of geometry with a fixed pool of active source frames, so it is bounded recall, not perfect recall, and the paper describes recovered geometry as recognizable rather than exact. Navigation and perspective remain the model's weakest dimensions. And the distilled checkpoints were trained on video-to-video conditioning only, so image-to-video and text-to-video are effectively untested capabilities on the model most people will download.
Originally published on Ground Truth, where every claim is checked against the primary source.
