Visko Orbis 1.0: Real-Time 4K Video Generation You Can Redirect Mid-Stream

Visko Orbis 1.0 does not finish a video and hand you a file. It starts streaming a scene the moment you give it a prompt and keeps generating — and when you type a new prompt, the scene changes around you mid-stream, no restart and no loading bar. Visko, the Sunnyvale startup behind it, calls this new class of system a "Live Model," and Orbis is its first one. The company came out of the gate on September 1, 2026 with public access at visko.ai, a technical report on arXiv, and a $10 million pre-seed round led by Llama Ventures. Orbis is a day old, so the benchmark figures below are Visko's own until someone independent runs them.

The reason to pay attention is not that Orbis is better than the video models you can already call — nobody has proven that yet. It is that Orbis is a different category. Every major video generator shipping today, from OpenAI's Sora 2 Pro to Google's Veo 3.1 to Kling 3.0, is what Visko calls a "query model": you send a prompt, wait for a render, and get a finished clip. Orbis runs instead of renders. It treats video as something that is always happening, and it lets you talk to it while it happens.

What "live" actually means

Visko's framing is that a query model has no internal clock and no memory, so the longer a generated sequence runs, the more it falls apart — artifacts, broken physics, color drift. Orbis is built to hold a world together. It maintains persistent internal state across the whole stream, so subjects, scenes, and style stay put from one chunk to the next. Visko's testing claims it can keep a scene coherent for an hour or more without the quality or color drift that long diffusion rollouts are known for.

Two capabilities fall out of that design. First, mid-stream prompt editing: you can add, change, or switch prompts while generation is running, and Visko says the output reflects the change in under a second on average — the world updates around it instead of restarting. Second, continuity: the model carries subjects, scenes, and style across chunks and sessions, which is what makes long-form generation possible at all.

What you can do with it right now

Orbis opens with four public modes, and the difference from a clip generator is visible in the list. "Choose What Happens Next" offers branching story choices mid-stream. "Speak It, See It" takes spoken narration and renders the scene live as you read. "Train and Test Robots" drops a robot into a generated world to practice before the real one. "Freeform Playground" takes text or an image-plus-text description and generates a world from scratch. Underneath sits a longer queue of example worlds — a virtual band, a kung fu courtyard, an astronaut working in orbit, a live-streaming mode — each a running scene you can steer rather than a rendered file.

• Output — 4K video at 24 FPS, streamed in real time (vendor-reported).

• Duration — hour-scale generation claimed, with no evident quality or color drift (vendor-reported).

• Inputs — long-form text-to-video, image-to-video, and video continuation; multilingual prompts; prompt-switching mid-generation.

• Interactivity — re-prompt at any moment; Visko reports updates in under a second on average.

• Memory — a bounded multi-scale memory preserves subjects, scenes, and style across chunks.

• Physics — a latent world model scores candidate futures at inference, steering output toward physically plausible motion.

How it works

Orbis is built on neural stochastic differential equations, or Neural SDEs — an architecture that lets the model evolve its output continuously over time instead of resolving one discrete clip. It runs on an internal "world clock" inside a unified latent space made of perception, memory, and physics tokens: a running account of what the model is seeing, what it remembers, and what the laws of motion should do next.

Generation is streaming, closer to next-token prediction in an LLM than to diffusion denoising. The model finishes the first frame, ships it, then generates the next frame conditioned on everything that came before. A distilled chunk-wise streaming generator handles the base stream while a streaming video upscaler raises it to 4K, all served by an optimized GPU serving engine that Visko says keeps up in real time.

Who is behind it

For a model with one day of public history, the team matters. Visko was founded in 2025 by Qing (Will) Yin, a Stanford Ph.D. in computational mathematics and mechanics who spent three years as a researcher at Apple. The 16-person team draws from Apple, Google DeepMind, Meta, Amazon, and Tesla, and the advisory board is chaired by UC Berkeley's Michael I. Jordan. That pedigree is real credibility behind the launch — and it is still not a substitute for an independent evaluation.

The numbers so far — and how much to trust them

Visko's technical report, "Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation" (arXiv: 2607.26694), makes a specific set of claims. On DOVER, an established metric for aesthetic and technical quality, Orbis 1.0 leads the compared real-time and long-video systems. On VideoAlign, a visual- and motion-quality metric, it ranks first as well. And in a long-form human-preference Arena study across eight systems, Visko reports the model took the highest overall-preference and temporal-stability ratings. The comparison set includes Odyssey's Odyssey-2, which Visko says Orbis beats on those benchmarks.

Keep the provenance straight: every one of those numbers is Visko's own, published the same day the model launched, and no independent lab or neutral leaderboard has run Orbis yet. To the report's credit, it also discloses dimensions where Orbis trails other systems — more than most launch-day reports do — but a self-evaluation is a starting point, not a score. There is no LMArena or Artificial Analysis entry for Orbis today. That is normal for a model that shipped yesterday, and it is exactly why the fair verdict right now is "promising, unproven."

The report above is the entire evidentiary record for Orbis's headline claims — the benchmark numbers live in that document, and the document is the vendor's own.

Access, cost, and the honest unknowns

You can try Orbis today, two ways. Visko has opened public access at visko.ai — the models page runs a live demo with a queue of example worlds, and the Freeform Playground takes text or image-plus-text input. The report's distribution note adds that Orbis is also available through Visko's partner Reactor, which offers both a dynamic version and a stable version, with a sandbox and an API. That API is the part developers care about: it is the first real route to building on the model rather than just watching it.

The models page above is the full public surface right now: demo cards, a playground, and no price tag anywhere.

What does not exist yet is a price. Visko has published no per-second, per-minute, or subscription pricing for Orbis, and neither its site nor the distribution partner lists one. The analyst reaction has flagged exactly this: Constellation Research's Holger Mueller called Visko's work "extremely encouraging" but noted that continuous dynamic generation "probably won't come cheap." The compute reality backs that up — streaming 4K at 24 FPS indefinitely is a very different cost curve from rendering a 10-second clip once.

For a day-old model with no price and no independent numbers, the sane way to try it is the same as for any unproven model: test traffic, not production path. When a provider lists Orbis, that is exactly what a routing layer is for — point a prompt at Orbis, keep a proven fallback on the same endpoint, and let automatic failover swap in the reliable model if the new one stalls or drifts. OrcaRouter runs one API across 200+ models with provider list prices passed through at 0% markup, so a price cut on any of them is live here the same day the vendor announces it. For the record: OrcaRouter does not host or route Visko Orbis 1.0 yet — it ships through Visko's own site and its distribution partner — and the pass-through math above is what applies the moment a provider lists it.

Who should care

Three audiences have a real reason to watch Orbis. Robotics and simulation teams first: the demo list includes training and testing robots in generated worlds, and Visko pitches Orbis as a way to produce training footage that would be unsafe or expensive to stage — corner cases for autonomous vehicles and humanoid robots. That is the same pitch a wave of world-model startups are making, and Orbis is early but built for it. Second, game and interactive-media developers, because a world that keeps generating and reacts to input in real time is a rendering backend, not a clip tool. Third, anyone running live AI video in front of a user — live commerce, virtual influencers, education — where steering the stream live is the product.

And the audience that should wait: anyone whose workflow is "render a clip, edit it, ship it." Orbis's streaming paradigm solves a different problem than batch video generation, and until pricing and API terms are public, the cost of an always-running stream is an unknown you should not underwrite.

What to watch next

Three things would turn this debut into a model you plan around. A first independent look — an LMArena or Artificial Analysis entry, or a third-party replication of the DOVER and VideoAlign claims. Published pricing, which will decide whether "runs rather than renders" is affordable for anyone beyond research teams and funded startups. And evidence at scale: Visko's own report is upfront that memory-structure design, detailed action descriptions in training data, and scaling remain open work, so the hour-scale claim deserves testing by someone who is not selling it.

Right now Visko Orbis 1.0 is the most interesting video model you cannot yet evaluate fairly — a genuinely new architecture, a day of public history, and an impressive set of self-reported numbers waiting for a second opinion.