Curious about today's AI digest?ai-tldr.dev

Daily Digest

Pomegra Startups

Visko Raises $10M to Stream 4K AI Video in Real Time

Visko Platform (US) — Stanford PhD and former Apple researcher exits stealth with $10M pre-seed from Llama Ventures, launching Orbis, a "live model" that generates hour-long 4K video at 24fps in real time and updates instantly when the user changes the prompt mid-stream.

FundingAINOTABLE4 min read
Visko Raises $10M to Stream 4K AI Video in Real Time

Sunnyvale startup Visko exits stealth with $10M from Llama Ventures, launching Orbis - a live model that streams 4K video at 24fps and updates instantly to mid-stream prompt changes.

Key Takeaways

  • Visko closed a $10M pre-seed round led by Llama Ventures on September 1, 2026
  • Orbis streams 4K video at 24fps continuously, sustaining hour-scale sessions without visual or color drift
  • Users can change prompts mid-generation; Orbis reflects edits instantly without pausing or restarting

Lead

Visko, a Sunnyvale, California AI research company, exited stealth on September 1, 2026, announcing a $10 million pre-seed round led by Llama Ventures and opening public access to Orbis, its first live model. Orbis streams interactive 4K video at 24 frames per second in real time - and keeps going for up to an hour without losing coherence. The round brings the 16-person company out of roughly a year of quiet development, with no valuation disclosed.

What Does Orbis Actually Do?

Orbis does not render and deliver a finished clip. It runs continuously, generating video frame by frame while the user watches, and accepts prompt changes at any moment during the stream. Most current AI video tools require a user to submit a prompt, wait - typically seconds to minutes - and receive a completed sequence that cannot be modified without restarting. Orbis eliminates that waiting and that finality. The model sustains hour-scale generation without evident visual drift or color inconsistency, a technical failure mode that has plagued earlier attempts at long-form AI video. Visko categorizes this approach as a "Live Model," distinguishing it from generation models that produce static outputs.

Underlying the system is physics-grounded generation and persistent memory across the session. That combination lets Orbis track spatial and object relationships over an extended sequence - so a scene does not subtly reshape itself every thirty seconds.

Who Built It, and From Where?

Visko was founded in 2025 by Qing (Will) Yin, who holds a Stanford PhD in computational mathematics and mechanics and spent three years as a researcher at Apple. The team of 16 draws from Apple, Google DeepMind, Meta, Amazon, and Tesla, a roster that reflects a deliberate bet on engineering depth over speed to market. The company is also working with Google, though the nature of that relationship has not been detailed.

Yin's academic background in computational mechanics is not incidental. Physics simulation - tracking how objects interact, how light scatters, how surfaces deform - has historically required purpose-built engines. Visko's claim is that a generative model can internalize enough of those constraints to produce consistent, grounded video over long time horizons without a separate physics engine running underneath.

Why Does the Prompt-Steering Capability Matter?

The ability to steer a generation mid-stream matters more for professional use cases than for one-off clip creation. A director or game designer working with an AI video tool today faces a binary: accept the output or restart. Mid-stream control turns the model into something closer to a collaborator - one that can be redirected without discarding what came before. For interactive applications, live events, or real-time content pipelines, that changes the production workflow structurally.

The category Visko is entering is genuinely contested. OpenAI, Google, Runway, and Kling have all advanced text-to-video capabilities materially in the past 18 months. None have publicly demonstrated stable hour-length generation with real-time prompt updates at 4K resolution. That gap - if Orbis delivers on its specifications at scale - is meaningful. Whether it translates into a durable moat depends on how quickly larger labs close it.

What Does the $10M Signal About the Round?

A $10 million pre-seed is large by historical standards for the stage but unremarkable in the current AI funding environment, where early rounds have inflated sharply since 2024. Llama Ventures, which focuses on pre-seed through Series A in AI and software, is the named lead. No co-investors have been disclosed, and the company has not stated a valuation. At this amount and stage, the round is structured to fund the team through a Series A rather than toward commercialization at scale.

The absence of revenue figures, customer commitments, or enterprise pricing is standard for a public launch paired with a pre-seed close. What Visko has done is sequence the announcement to coincide with open public access - which means Orbis is testable now, not just described.

Outlook

Visko enters a crowded video-generation market with a technically distinct claim: continuous, steerable, hour-length 4K output. The $10M pre-seed gives it runway to prove the technology holds at scale and to build toward a Series A. The near-term question is whether public access to Orbis generates the user and developer traction that justifies a larger raise - and whether the core technical advantages survive contact with a field moving as fast as it currently is.

More Startup News