Visko, a Sunnyvale startup barely a year old, pulled in $10 million in pre-seed funding on Monday and opened public access to what it claims is a breakthrough: a foundation model that generates hour-long 4K video in real time. Llama Ventures led the round, though the company declined to discuss valuation.
The funding figure stands out. It's five to ten times typical pre-seed medians — which range from around $1 million for pre-seed alone to $2 million for combined pre-seed and seed rounds in early 2026, depending on the dataset. Visko says the capital will push forward research, grow its 16-person team and keep developer access to the model open.
That model, Orbis 1.0, produces 4K footage at 24 frames per second as it runs. The company says generation can stretch to an hour, and users can change prompts mid-stream without restarting from scratch. Visko describes the system as "physics-grounded" with persistent memory — terms that suggest the model maintains spatial and narrative consistency across long sequences. A technical paper posted to arXiv on July 29 outlines the architecture: a chunk-wise streaming generator paired with a streaming upscaler.
"We built Orbis to hold together — physically, visually and narratively — for as long as you want to stay inside it," founder and CEO Qing Yin said in the announcement. Whether that holds true under real-world use remains to be seen.
Distribution runs through Reactor, a platform that left stealth in June with $59 million to build infrastructure for what it calls real-time AI worlds. Reactor's API documentation lists two Orbis endpoints: Visko Orbis Dynamic, which supports live resolution switching, and Visko Orbis Stable. The partnership suggests Visko is betting on platform distribution rather than building its own infrastructure from the ground up.
Yin earned a Stanford PhD in computational mathematics and mechanics in 2021. She worked as a researcher at Apple before launching Visko in 2025. Michael I. Jordan, the UC Berkeley professor whose work underpins much of modern machine learning, chairs the startup's advisory board. Steve WaiChing Sun of Columbia and Mengye Ren of NYU round out the advisors.
"Visko delivers a state-of-the-art architecture that has a meaningful technological lead beyond conventional short-clip video generation," Jordan said in the announcement. His endorsement carries weight, though distinguishing genuine technical leaps from incremental improvements in this space has grown harder as competition intensifies.
And the competition is real. Odyssey-1 positioned itself in May 2025 as a real-time, playable world model. Decart's Oasis 3 targets real-time multi-view simulations for autonomous systems. World Labs released Marble, a multimodal world model, earlier this year. Each system navigates different tradeoffs between resolution, speed, duration and interactivity. Some prioritize rendering speed. Others optimize for visual fidelity or interactive control. Visko's pitch centers on duration and consistency, but whether that proves more commercially valuable is an open question.
The company's announcement did not disclose named commercial customers, pricing tiers or usage limits. The company is incorporated in Delaware and registered in California at 880 West Maude Avenue in Sunnyvale. For now, the startup appears focused on proving the technology works at scale before locking in a business model. That's a familiar path for foundation model companies, though not all survive the transition from technical demonstration to sustainable revenue.
