Visko's Orbis 1.0 is a foundation model for real-time interactive video, aimed at researchers, game and simulation teams, robotics groups, and creative tools builders who need generated worlds that keep running rather than one-shot clips. It supports long-form text-to-video, image-to-video, video-to-video continuation, multilingual prompting, and in-generation prompt switching , so a user can drop in a photo, describe a scene, and then steer what happens next while the stream continues. The launch matters because conventional AI video generates a clip and stops, whereas Orbis generates video continuously and streams it live, with users able to change the prompt mid-generation and see the update reflected instantly .
Under the hood, Visko frames Live Models as a new class of world models that run rather than render . A latent-video model generates successive temporal chunks under a single conditional formulation, with each chunk reading the active event condition, optional visual references, and a bounded multi-scale memory of the preceding rollout, so the persistent visual state of a session is separated from the user's transient intent . A prompt update becomes visible in under one second on average, and decoded chunks are delivered progressively rather than after the full video completes . Built on a distilled chunk-wise streaming generator and a streaming video upscaler, Orbis targets 4K video generation at 24 FPS in real time using an optimized GPU serving engine , with a bounded multi-scale memory that preserves subjects, scenes, and style across chunks to sustain hour-scale rollouts without evident quality or color drift .
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 , and is based in Sunnyvale, California. The 16-person team draws from Apple, Google DeepMind, Meta, Amazon, and Tesla , and is advised by Michael I. Jordan of the University of California, Berkeley, Steve WaiChing Sun of Columbia University, and academic partner Mengye Ren of New York University . The company opened public access to Orbis alongside a $10 million pre-seed round led by Llama Ventures , with API distribution handled through partner Reactor.
Comments (11)
Sign in to join the discussion.
P
Priya ThankachanSep 1, 2026
The demo reel is gorgeous but I clocked the cut every 8 seconds. what happens at minute 3 when the physics has to remember what a chair is?
K
Kwame OyelaranSep 1, 2026
"Live Models to predict the world" is doing a lot of heavy lifting. try: "worlds that keep going after you look away." free of charge, tip jar in bio.
S
Sunna OttoSep 1, 2026
the launch page loads straight into a playable world with no signup wall. that is the flex, not the tweet copy.
D
Dmitri V.Sep 1, 2026
ok wait, streaming a persistent world in real time with unbounded length is either the coolest thing this year or a very expensive screensaver. either way I'm clicking.
R
Renata ColomboSep 1, 2026
before I say interesting, what does session length actually look like and how are you thinking about compute per active world. "unbounded" is a fun word for a finance team.
H
Hiro TanabeSep 1, 2026
we prototyped something like this on the Stadia team in 2019 and it died in a room with no windows. rooting for you to actually ship it this time.
M
Malik FerreiraSep 1, 2026
orbis 1.0, or as my GPU calls it, the eviction notice. jokes aside the persistence claim is the only one that matters, please don't let me down.
A
Anneliese BraunSep 1, 2026
the tweet buried the API mention in the second line. that's the actual story and you're treating it like a footnote.
R
Ravi DeshmukhSep 1, 2026
world models are a category in search of a customer. game studios don't want your physics and film people don't want your latency. convince me otherwise.
T
Tolu AdebayoSep 1, 2026
reminds me of the early Runway and Decart bets in my book, but with actual state across frames. betting on the memory piece is the right wedge.
Z
Zee NakamuraSep 1, 2026
building an adjacent thing in the sim space and honestly the persistent memory demo made me reconsider a few architectural choices tonight. nice work team.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Visko
We build Live Models to predict the world.
We build Live Models to predict the world.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Visko's Orbis 1.0 is a foundation model for real-time interactive video, aimed at researchers, game and simulation teams, robotics groups, and creative tools builders who need generated worlds that keep running rather than one-shot clips. It supports long-form text-to-video, image-to-video, video-to-video continuation, multilingual prompting, and in-generation prompt switching , so a user can drop in a photo, describe a scene, and then steer what happens next while the stream continues. The launch matters because conventional AI video generates a clip and stops, whereas Orbis generates video continuously and streams it live, with users able to change the prompt mid-generation and see the update reflected instantly . Under the hood, Visko frames Live Models as a new class of world models that run rather than render . A latent-video model generates successive temporal chunks under a single conditional formulation, with each chunk reading the active event condition, optional visual references, and a bounded multi-scale memory of the preceding rollout, so the persistent visual state of a session is separated from the user's transient intent . A prompt update becomes visible in under one second on average, and decoded chunks are delivered progressively rather than after the full video completes . Built on a distilled chunk-wise streaming generator and a streaming video upscaler, Orbis targets 4K video generation at 24 FPS in real time using an optimized GPU serving engine , with a bounded multi-scale memory that preserves subjects, scenes, and style across chunks to sustain hour-scale rollouts without evident quality or color drift . 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 , and is based in Sunnyvale, California. The 16-person team draws from Apple, Google DeepMind, Meta, Amazon, and Tesla , and is advised by Michael I. Jordan of the University of California, Berkeley, Steve WaiChing Sun of Columbia University, and academic partner Mengye Ren of New York University . The company opened public access to Orbis alongside a $10 million pre-seed round led by Llama Ventures , with API distribution handled through partner Reactor.
Comments (11)
The demo reel is gorgeous but I clocked the cut every 8 seconds. what happens at minute 3 when the physics has to remember what a chair is?
"Live Models to predict the world" is doing a lot of heavy lifting. try: "worlds that keep going after you look away." free of charge, tip jar in bio.
the launch page loads straight into a playable world with no signup wall. that is the flex, not the tweet copy.
ok wait, streaming a persistent world in real time with unbounded length is either the coolest thing this year or a very expensive screensaver. either way I'm clicking.
before I say interesting, what does session length actually look like and how are you thinking about compute per active world. "unbounded" is a fun word for a finance team.
we prototyped something like this on the Stadia team in 2019 and it died in a room with no windows. rooting for you to actually ship it this time.
orbis 1.0, or as my GPU calls it, the eviction notice. jokes aside the persistence claim is the only one that matters, please don't let me down.
the tweet buried the API mention in the second line. that's the actual story and you're treating it like a footnote.
world models are a category in search of a customer. game studios don't want your physics and film people don't want your latency. convince me otherwise.
reminds me of the early Runway and Decart bets in my book, but with actual state across frames. betting on the memory piece is the right wedge.
building an adjacent thing in the sim space and honestly the persistent memory demo made me reconsider a few architectural choices tonight. nice work team.