Qing Yin's Visko raises $10M for AI worlds that keep running

Why it matters

Visko is trying to move generative video from finished clips into continuously running environments. If Yin can make that affordable and reliable, Orbis could serve robotics and simulation as well as media.

Qing (Will) Yin, Visko's founder and CEO, led the Sunnyvale, California AI research company through a $10M pre-seed round and the launch of Orbis on September 1st. The financing backs Yin's thesis that generative models should run continuously instead of producing a clip and stopping.

Llama Ventures led the round, according to FinSMEs' financing report and Visko's announcement. Visko says it was founded in 2025 and employs 16 people drawn from Apple, Google DeepMind, Meta, Amazon and Tesla.

Yin arrived at video generation through the mechanics of the physical world. He earned a bachelor's degree in civil engineering from Tianjin University in 2014 and completed a Stanford Ph.D. in civil and environmental engineering in 2020. His doctoral research examined the long-term, time-dependent behavior of geomaterials across different spatial and temporal scales. He later worked on data-driven plasticity at Columbia University before joining Apple as a research and development engineer.

That background is unusually direct preparation for Visko's central problem: keeping a generated environment physically and visually coherent as time passes and users intervene.

A model designed to stay on

In an essay introducing Visko's Live Models, Yin and his co-authors described the next AI paradigm as "a process that runs, in step with the world itself."

Orbis is Visko's first attempt to turn that idea into a product. Users can begin with text, an image or an existing video, then change prompts while generation continues. Visko says Orbis incorporates those instructions without restarting the stream, while persistent memory carries subjects, scenes and visual style across successive chunks.

The distinction is consequential for applications that require a feedback loop. A pre-rendered video can illustrate a scene. A continuously generated environment could respond to a robot's actions, let a player alter a simulated world or allow an operator to change conditions during a training exercise.

Persistent memory is meant to prevent characters, objects and scenery from gradually changing identity during long sessions, a familiar failure mode in generative video.

Visko's Orbis 1.0 technical paper, posted on July 29th, claims real-time 4K output at 24 frames per second and hour-scale generation. Those results come from Visko's own research and have not been independently benchmarked. The paper's claims do not settle how Orbis will perform under sustained customer workloads.

Public launch

Visko's September 1st announcement says Visko opened public access to Orbis.

Pricing, usage limits and infrastructure requirements are outside the launch announcement. Those figures will matter because real-time video generation can be expensive to serve.

Visko has named Reactor as a distribution partner, but has not detailed whether Orbis will be sold through an API, a hosted application or enterprise agreements. The announcement also does not identify paying customers or commercial pilots. For a research-heavy company at the pre-seed stage, the immediate job is converting a technically ambitious demonstration into something developers can afford and reliably integrate.

A crowded definition of world models

The phrase "world model" already covers products with different architectures and customers. Odyssey develops streaming, playable environments. Decart's Oasis 3 targets real-time, multi-view simulations for autonomous vehicles and other physical AI systems. World Labs' Marble creates editable 3D environments that can be exported as meshes, Gaussian splats or video.

Visko is positioning Orbis around continuous video, mid-stream prompt changes, persistent memory and physical plausibility. That places Visko between creative video generation and closed-loop simulation, with potential uses in gaming, robotics, education, live commerce and media. Each of those markets has different tolerance for latency, visual errors and serving costs, so Visko will eventually need a narrower commercial entry point than the launch-day list.

Capital has moved quickly into the category. World Labs announced $1B in new funding on February 18th, while Runway raised a $315M Series E on February 10th to expand its work on world simulation. Those later-stage financings show the scale of capital already moving into the category, though they are not direct comparisons with Visko's pre-seed.

Visko says Michael I. Jordan of the University of California, Berkeley chairs its advisory board, with Columbia's Steve WaiChing Sun and New York University's Mengye Ren serving as advisers or academic collaborators. Visko plans to use the financing to continue developing Orbis and expand its team.

Yin has translated a career in computational mechanics into a clear product thesis: generated worlds should preserve state, obey motion and accept intervention while they run. The $10M gives Visko room to prove that thesis beyond its own paper. Customers will judge Orbis on latency, cost and control under sustained use, measurements that a launch-day benchmark cannot settle.