Orbis Lets Users Alter Prompts Mid-Stream as Visko Takes Its Video Model Public

Visko has opened public access to Orbis 1.0, a video foundation model built to continue streaming while users change prompts. The launch, alongside a $10 million pre-seed round led by Llama Ventures, puts the startup’s central proposition in users’ hands: generated video as an interactive process rather than a one-off result.

One session, not a sequence of restarts

Visko characterizes most AI video tools as offline systems: a person submits a prompt, waits for a completed clip, then begins another generation to make a change. Orbis instead continuously streams video. Visko says prompt edits are reflected during generation while the stream continues, and that the model maintains persistent memory.

A company-defined category with a physics claim

Visko calls this design a Live Model, its term for a foundation model intended for continuous, real-time world intelligence. The company contrasts it with models that accept an input, return a prediction, and stop. Its technical report says Orbis uses a latent world model to score candidate futures during inference, steering output toward physically plausible motion and object dynamics.

Visko says Orbis generates 4K video at 24 frames per second in real time and can sustain hour-scale output without evident visual or color drift. Those are company-reported performance claims, but they identify the difficult product requirement behind the pitch: an editable stream must remain coherent as it runs.

Benchmarks offer an initial, limited test

In its technical report, Visko says Orbis led the systems it compared on DOVER aesthetic and technical quality, VideoAlign visual and motion quality, and overall-preference and temporal-stability ratings in a long-form human-preference study. The report also identifies dimensions where Orbis trails other systems. The results are Visko’s own evaluation, not an independent ranking.

What public access now tests

• Whether mid-stream prompt changes remain responsive while output continues.
• Whether persistent memory and long-running visual consistency hold outside Visko’s reported evaluation setup.

Visko was founded in 2025 and is based in Sunnyvale, California. Founder Qing (Will) Yin previously spent three years as an Apple researcher. The new financing gives the young company backing as it moves its interaction-first video thesis from a technical report into public use.