Visko Raises $10M for Real-Time AI Model Orbis

Visko has raised a $10M Pre-Seed round led by Llama Ventures and opened public access to Orbis, the company's first Live Model. The Sunnyvale startup says Orbis can generate interactive 4K video at 24 frames per second while retaining enough state to keep a scene coherent as a user changes the prompt.

The announcement puts capital behind a different product assumption for generative video. Most video models return a finished clip after receiving a request. Visko wants Orbis to keep running, remember what has already happened, and respond while the world is still being generated.

That distinction reaches beyond video quality. If a model becomes a persistent environment for games, education, robotics, simulation, or digital companions, customers are no longer buying only a better clip. They are buying continuity, and continuity turns memory, inference cost, reliability, and control into product requirements.

What Visko Announced

Visko disclosed the financing and Orbis launch on September 1, 2026. The $10M Pre-Seed round was led by Llama Ventures, but Visko did not name other investors, disclose a valuation, or provide a detailed allocation for the capital. Those omissions matter at this stage because the announcement combines a financing event with a public product release, while leaving the commercial plan largely outside the frame.

Visko was founded in 2025 by Qing (Will) Yin, the company's founder and CEO. The company reports a 16-person team with experience from Apple, Google DeepMind, Meta, Amazon, and Tesla. Its advisers include Michael I. Jordan of the University of California, Berkeley and Steve WaiChing Sun of Columbia University.

Orbis is available through Reactor, Visko's product interface. The company positions the system for entertainment, education, robotics, and simulation, although it has not disclosed customer names, revenue, pricing, or production deployments. Public access gives outside users something more useful than a private demo, but it does not yet answer whether long-running generation can become dependable infrastructure.

How Orbis Changes the Video Product

Conventional generative-video systems usually behave like query engines. A user submits text, an image, or another input; the model processes that request; and the system returns a bounded video. Editing often means generating again, which can change the character, camera, room, color, or sequence in ways the user never requested.

Visko describes Live Models as continuously running systems that remain responsive to new input. Orbis accepts prompt changes during generation and carries the scene forward instead of treating every intervention as a new job. The company's product materials say users can redirect action, perspective, or style while preserving the broader state of the world.

The idea sounds subtle until the user expects the scene to remember something. A generated doorway should remain where it was placed. A character should not quietly change identity. An object moved 30 seconds ago should not reappear because the model lost the plot. Video AI has become good at making moments; a live environment has to survive consequences.

What the Technical Evidence Shows

Visko's Orbis technical report describes a latent world model, persistent multi-scale memory, streaming super-resolution, and a mechanism for evaluating possible futures for physical plausibility. Visko reports interactive 4K output at 24 frames per second and hour-scale generation without evident degradation in visual quality or color stability.

The report also presents human-preference, temporal-stability, DOVER, and VideoAlign comparisons against other video systems. Orbis ranks first overall in several company-authored evaluations, while the detailed tables show that it does not lead every measure. That is a healthier reading than turning a benchmark table into a parade float.

The evidence boundary is important. Visko and its research team produced the report, selected the testing conditions, and published the results. Independent coverage from SiliconANGLE and TestingCatalog confirms the launch and core product claims, but it does not provide an independent technical replication. The next proof will come from outside users reproducing stability, latency, controllability, and cost under real workloads.

Why Continuous Generation Changes the Economics

A clip has a natural stopping point. A world designed to remain available keeps consuming memory, inference, and operational attention for as long as somebody expects it to stay alive. That turns the world-hour, rather than the isolated generation, into a possible unit of value.

The change creates a new set of product questions. Developers need to know how state survives longer sessions, how much intervention the model can absorb before continuity weakens, and how infrastructure cost grows with resolution, frame rate, and memory. Customers need pricing they can understand before an ambitious simulation becomes the cloud bill wearing a costume.

Those economics could separate impressive research from durable software. A game studio may tolerate occasional visual imperfections if the environment is responsive and cheap enough to operate. A robotics team may care less about cinematic beauty than causal consistency. An education product may need a persistent scene to remain safe, controllable, and accurate across repeated student interactions.

The Market Around Live Models

The broader AI-video market has spent years competing on resolution, duration, motion quality, prompt adherence, and time to render. Visko is pushing the competition toward persistence and interaction, where the output is not simply watched but inhabited, redirected, and expected to remember.

That places Orbis near the emerging market for world models and causal AI. The categories are not identical, but they share a demand for systems that represent changing state rather than predict a single polished answer. Generating a plausible next frame is useful. Maintaining a world whose next frame respects everything that came before is a much larger promise.

Visko's team composition and advisers suggest the company understands that the problem sits between research and systems engineering. The model has to generate; memory has to persist; the interface has to accept intervention; and the infrastructure has to deliver the experience without turning latency into the main character.

What the $10M Makes Possible

The Pre-Seed round gives Visko time to turn its public release into an operating discipline. The company has not specified how the capital will be divided among research, computing infrastructure, hiring, or commercialization, so any detailed use-of-funds claim would be invented. What is visible is the sequence: Visko has paired an unusually large early round with a product that outsiders can begin testing immediately.

For investors and operators, the central question is not whether Orbis can produce a striking demonstration. The internet has a healthy surplus of striking demonstrations. The question is whether continuous generation can remain coherent, responsive, affordable, and useful long enough for a customer to build a recurring habit around it.

If Visko can answer that question, the company will not merely have extended the length of an AI video. It will have changed what the user believes the model is doing. The clip disappears into the feed; the world has to be there when somebody comes back.