Visko Orbis keeps the world running

A live model for interactive long video. Change the prompt. The stream stays on.

Visko / Llama VenturesOrbis 1.0 technical report4K / 24 FPSHour-scale memory

The product lets you steer a world

Most video models take a prompt and return a clip. Visko Orbis turns generation into a session that is still happening.

Text to world

Text to world

Describe a place, a person, and what happens next. Orbis starts a stream instead of a finished clip.

Image to world

Image to world

Drop a still. The model holds identity, wardrobe, and room while motion begins.

Keep a clip going

Keep a clip going

Feed a video and the same world continues. Memory stays bounded so hour-scale rollouts do not drift off color.

Switch the prompt live

Switch the prompt live

Type a new event while the stream is running. The paper reports the change becomes visible in under one second.

Worlds already in motion

Hover to play. These clips follow the official demo worlds: orbit work, courtyard combat, a bay meteor, and robot practice.

Astronaut

Astronaut

Fox companion

Fox companion

Kung fu courtyard

Kung fu courtyard

Virtual band

Virtual band

Bay meteor

Bay meteor

Robot practice

Robot practice

Elf archer

Elf archer

Voxel forest

Voxel forest

How Orbis 1.0 holds a world

The model splits persistent visual state from the prompt you just typed. Each chunk reads the active event, optional references, and a bounded memory of what already happened.

Real-time generation

4K at 24 FPS through a distilled chunk generator and a streaming upscaler.

No fixed length

The session runs until you stop it. Worlds are not capped at a few seconds.

Persistent memory

A bounded multi-scale memory keeps subjects, rooms, and style across chunks.

Interactive

Prompt switching, voice events, and later expert models share one future process.

Physics-grounded

A latent world model scores candidate futures and steers motion toward plausible dynamics.

Full-duplex

You speak or type while the stream is already moving. Input and output overlap.

Rewrite the next beat. The world stays live.

This is the in-generation prompt switch. Pick an event and the stream cuts to that world. Open the live demo when you want a full session.

An astronaut works outside a space station. Earth hangs in the distance. A tool drifts, then the visor turns toward a sudden warning light.

A session from prompt to stream

  1. Write the beat

    Text, a still, or an existing clip can start the same future process.

  2. Steer while it runs

    A new prompt becomes the active event. The stream does not stop to re-render from zero.

  3. Stay in the same world

    Memory keeps people and light across chunks. When credits run out, add more on the pricing page.

Questions

Is this the official Visko site?

No. This is an independent playground for Visko Orbis. Research, papers, and the first-party demo live at visko.ai.

What can the product do today?

Long-form text-to-video, image-to-video, video continuation, multilingual prompts, and prompt updates while generation is running.

What is special about the Orbis 1.0 model?

It is a Live Model. It streams chunks with persistent memory, reports 4K at 24 FPS, and keeps hour-scale rollouts from obvious quality or color drift.

Why are sessions only 30 seconds?

The official model runs on Reactor. A GPU is billed every second it is held. We prepaid a 30-second block and cut the session when time is up, so an idle tab cannot keep spending.

How do credits work?

No free trial. 1 credit is 1 second of Orbis Stable. A 30-second session costs 30 credits. Dynamic costs 3 credits per second. Pricing defaults to yearly Orbit at $29/mo, billed $348.

Open a live session now

Buy credits first. Official Orbis on Reactor bills by the second. We prepaid 30 seconds.