Visko Platform Inc., an artificial intelligence research lab focused on developing persistent world models, said today it has raised $10 million in a pre-seed funding round from Llama Ventures.
It’s also opening public access to its first foundation model, Orbis, which is based on a new paradigm the company calls “live models.”
The startup says it is trying to solve one of the most annoying problems with existing AI video generation models and world models, which have a tendency to break down thanks to their inability to maintain logical and visual coherence over extended durations. Conventional AI video generators such as Sora are built on “query models,” where users enter a prompt, wait and then receive the output in the form of a short, static, offline video clip.
In contrast, Visko’s live models are designed to run continuously. Orbis continuously generates and streams interactive visual worlds in real time, as opposed to rendering static files, which means that users can intervene at any moment. For instance, they can adjust or add a new prompt that instructs the character in the video to do something different, and the output will immediately reflect that request, without stopping the video stream.
Visko explains in a blog post that its live models are built on a completely different architecture to query models. The latter generates content in isolation, essentially on a clip-by-clip basis with no sense of time or persistent memory. This means query models have to rebuild the world’s state each time a new prompt or input is received. Over time, they suffer from temporal degradation that leads to visual artifacts, broken physics and color drift.
Orbis works differently. It generates video in a continuous, learned dynamic system that’s governed by neural stochastic differential equations or Neural SDEs. What this means is that instead of treating each frame as a discrete output call, Orbis runs on an internal “world clock” within a unified latent space that’s built from perception, memory and physics tokens. This architecture allows Orbis to maintain a persistent internal state that’s capable of anticipating continuous motion, retaining object detail and performance, and absorbing any real-time interventions without impacting the visual coherence of its outputs, even over very long time frames.
Founder and Chief Executive Qing Yin said long-term coherence has been the biggest challenge in AI video generation, because the longer a generated world runs, the more it falls apart. “We built Orbis to hold together, physically, visually and narratively, for as long as you want to stay inside it,” he said.
Orbis, which is available for anyone to play with on Visko’s website, streams its interactive video at 4K resolution at 24 frames per second. Yin said it can continue generating worlds for hours without any observable degradation in color or quality.
To back up these claims, it published a technical report that evaluates Orbis alongside a number of competing real-time and long-video systems on the DOVER aesthetic and VideoAlign visual and motion quality benchmarks. It beat out all of its competitors, including Odyssey’s Odyssey-2, and also ranked first in a human preference Arena study that compared the different model’s long-form performance.
Constellation Research analyst Holger Mueller said Visko’s work is extremely encouraging, because there’s so much potential in being able to accurately simulate real-world environments without them falling apart.
“Visko’s work shows that this goal now seems achievable, especially with such an abundance of compute resources now available,” he said. “Visko generates its world models dynamically rather than statically, so in theory they never have to stop. This is a key change of paradigm and should get a lot of interest, assuming its claims hold up in the real world. Cost may be an issue too, for dynamic, continuous generation probably won’t come cheap, but for now it’s great to see the innovation.”
Visko is being advised by the University of California at Berkeley professor Dr. Michael I. Jordan, who said that the startup is trying to solve one of the most significant problems with real-time AI video generation. “Visko delivers a state-of-the-art architecture that has a meaningful technological lead beyond conventional short-clip video generation,” he said. “The same core capability could support a remarkably broad range of commercial opportunities, from robotics and physical simulation to gaming and real-time creative media.”
