Yeta Labs announced a demo of a massive causal video model built with a hybrid stack of Ascend NPUs and NVIDIA GPUs, and opened a waitlist for access at https://world.yeta.ai. The tweet @Fried_rice shows a short clip generated by the system.
What the model can generate
The model produces video sequences where each frame is conditioned on previous ones, allowing coherent motion and scene changes from a single seed frame or text prompt. Early demos suggest it can handle simple actions and background continuity better than frame‑by‑frame generators.
Training hardware mix
Yeta’s engineers combined Ascend NPUs—optimized for large matrix ops—with NVIDIA GPUs for CUDA‑based kernels. This hybrid approach lets them offload specific tensor calculations to the NPUs while keeping the rest of the pipeline on familiar GPU tooling. The exact compute budget wasn’t disclosed, but the term “massive” implies many petaflop‑days of training.
Cost, access, and pricing unknown
Access is currently limited to a waitlist, and no pricing details have been published. Expect high inference costs given the model’s size and the need for both NPU and GPU resources. Startups should budget for cloud‑based NPU instances or consider on‑premise hardware if they plan to run the model at scale.
When to try it and what to watch
If your product needs short, coherent video clips (e.g., UI demos, marketing snippets), sign up for the waitlist and run a small proof‑of‑concept once you gain access. Keep an eye on latency and cost per generated minute; early reports hint at noticeable compute overhead. Also watch for potential bias or artifact patterns that can appear in generative video pipelines.