Visko Platform Inc., an artificial intelligence research lab, today announced a $10 million pre-seed funding round and the public debut of Orbis, its first foundation model for generating long-form, real-time interactive videos. The company, backed by venture capital firm Llama Ventures, is betting on a new paradigm it calls "live models" to overcome the limitations of conventional AI video generators.
The funding round, led by Llama Ventures, arrives as Visko opens public access to Orbis on its website. The model streams interactive video at 4K resolution and 24 frames per second, a technical leap over the short, static clips produced by existing tools. Visko's founder and CEO, Qing Yin, framed the launch as a response to a persistent problem in the field: keeping generated worlds coherent over time.
"We built Orbis to hold together, physically, visually and narratively, for as long as you want to stay inside it," Yin said.
The Limits of Query Models
Conventional AI video generators like Sora are built on what Visko calls "query models." These systems generate short, static, offline video clips after a prompt, producing content in isolation, clip-by-clip, with no sense of time or persistent memory. According to Visko, query models suffer from temporal degradation, leading to visual artifacts, broken physics, and color drift as sequences lengthen.
Existing AI video generators often break down entirely because they cannot maintain coherence over long durations. A query model treats each request as a fresh problem, with no internal clock or memory of what came before. The result, Visko argues, is video that looks convincing for a few seconds but falls apart under sustained scrutiny.
Orbis takes a different approach. It runs continuously and streams interactive visual worlds in real time. Users can intervene at any moment with new prompts, and the output changes immediately without stopping the stream. That capability, Visko says, comes from an architecture built on neural stochastic differential equations, or Neural SDEs, which allow the model to evolve its output smoothly over time.
Inside Orbis's Architecture
Visko published a blog post explaining the architecture behind Orbis. The model runs on an internal "world clock" within a unified latent space built from perception, memory, and physics tokens. That design lets Orbis maintain a persistent internal state for continuous motion, object detail, and real-time interventions.
The technical foundation is significant. Neural SDEs model the evolution of a system over time with a stochastic component, which suits the task of generating video frames that flow naturally from one to the next. The world clock provides a temporal anchor, so the model knows where it is in a sequence and can keep objects, lighting, and physics consistent.
Visko also published a technical report evaluating Orbis on DOVER aesthetic and VideoAlign visual and motion quality benchmarks. The company claims Orbis beat all competitors, including Odyssey's Odyssey-2, in those tests. Orbis also ranked first in a human preference Arena study for long-form performance.
Yin claims Orbis can generate worlds for hours without observable degradation. That claim, if accurate, would mark a major departure from the clip-based generation of query models, which typically lose coherence within seconds or minutes.
Expert Reactions and Caution
Holger Mueller, an analyst at Constellation Research, called Visko's work "extremely encouraging" and said it shows the goal of simulating real-world environments is achievable. He noted that the shift to dynamic generation is a "key change of paradigm," but he also sounded a note of caution.
"Visko's work shows that this goal now seems achievable, especially with such an abundance of compute resources now available," Mueller said.
Mueller added that "assuming its claims hold up in the real world" and that "cost may be an issue too, for dynamic, continuous generation probably won't come cheap." Running a model that streams 4K video at 24 frames per second, indefinitely, requires far more compute than generating a single clip. That cost could limit adoption, at least initially.
Visko is advised by Dr. Michael I. Jordan, a professor at the University of California at Berkeley. Jordan praised the company's technical direction and its commercial potential.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
"Visko delivers a state-of-the-art architecture that has a meaningful technological lead beyond conventional short-clip video generation," Jordan said.
Jordan also pointed to applications beyond entertainment. "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," he said.
A New Paradigm Takes Shape
The launch of Orbis comes at a time when the AI video generation market is crowded with players, but most are still shipping query-based tools. Visko's live models paradigm represents a structural shift, moving from generating discrete outputs to maintaining a persistent, evolving world.
For users, the difference is tangible. With a query model, you type a prompt and wait for a clip. With Orbis, you enter a world that keeps running, and you can change its direction at any moment. That interactivity opens new possibilities for gaming, virtual environments, and real-time creative work, though Visko has not yet detailed pricing or commercial plans.
The company's technical report and benchmark results are public, but independent verification of its claims has not yet been published. Mueller's caution about real-world performance and cost reflects the gap between a controlled benchmark and production use.
Visko's funding round, at $10 million, is modest by AI industry standards, but the company's backers are betting on the architecture rather than the scale. Llama Ventures led the round, and the funds will presumably support further development of Orbis and the underlying live models research.
The timing is notable. As of September 01, 2026, updated at 09:00 EDT, Visko has moved from research lab to public-facing product company. The article announcing the launch also mentioned other recent stories in the AI space, including work from Orchestra, Filigran, DataAgent, the Bank of England, OpenAI, and CrowdStrike, suggesting a busy period for the sector.
What Comes Next for Live Models
The live models paradigm is still young. Visko has published its architecture and benchmarks, but the broader research community will need to replicate and stress-test the results. The company's claims about hours-long coherence are bold, and independent evaluation will determine whether they hold.
Mueller's point about compute costs is central. Continuous generation at 4K resolution and 24 frames per second is computationally expensive, and the economics of running such a service at scale are unclear. Visko has not disclosed its infrastructure costs or pricing model.
Still, the direction is clear. If live models deliver on their promise, they could replace the clip-based workflow that dominates AI video today. The ability to intervene in real time, without stopping the stream, changes what creators can do with generative video.
Jordan's mention of robotics and physical simulation points to a broader thesis. A model that can maintain a persistent, physically coherent world could be used to train robots, simulate environments, or power interactive media. Those applications are speculative for now, but they explain why a professor of Jordan's stature would advise the company.
Visko's launch is a statement of intent. The company is not just releasing another video generator; it is proposing a new way to think about generative models. Whether the market embraces that paradigm will depend on real-world performance, cost, and the quality of the experience Orbis delivers.
For now, the public can try Orbis on Visko's website and judge for themselves. The company has made its model accessible, its benchmarks public, and its claims clear. The next step is verification.