---
title: "Tencent's Open World Model: Can an RTX 4090 Run It?"
url: "https://goremy.ai/blog/articles/tencent-open-world-model-run-locally-rtx-4090"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-20T05:40:17.306Z"
updated: "2026-11-20T05:40:17.306Z"
topic: AI Image, Video & Audio
description: "Tencent's open world model builds a consistent video world in 5-second chunks. Here's what it does, what hardware it needs, and how it held up in testing."
---

# Tencent's Open World Model: Can an RTX 4090 Run It?

> Tencent's open world model builds a consistent video world in 5-second chunks. Here's what it does, what hardware it needs, and how it held up in testing.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-20

## What is Tencent's open world model?

Tencent's open world model is an AI video system that generates a world you can keep directing, instead of a single clip you have to restart. You give it a scene and a subject, then keep adding instructions: add a dog, make it snow, swing the camera left. It builds the video in 5-second pieces and remembers what it already drew, so the world stays the same when you look away and back. The code and model are free, including for commercial use.

## TL;DR

- Tencent's model generates video in **5-second segments**, and each new segment is conditioned on everything generated before it, which is how it keeps the world consistent.
- You can steer it with **text, camera moves, or both**, and a session can run for about a minute.
- It has **no game engine** underneath. The video model does the remembering on its own.
- Tencent says it runs on **a single RTX 4090**, but that is a compressed, slower version. The real-time setup Tencent tested used **eight data center GPUs**.
- The code and model are **open and usable commercially**, but Tencent hasn't published comparisons against other world models, so the consistency claims come from its own demos.
- In one creator's hands-on test, the output was **wonky and not production quality**, though it looks useful for rough shot planning.

## How does the model keep a world consistent?

Most AI video tools work like a vending machine. You write one prompt, you get one clip, and if you want a change you start over. A world model works differently. The video keeps going, and you keep giving direction.

Tencent's model does this in 5-second pieces. Each new piece takes your latest direction, which can be text, a camera move, or both. Before it draws anything, it looks back at everything it has already produced. That memory is what stops the world from quietly rewriting itself.

You can see the problem it solves in a demo clip: an astronaut looks around, turns away, then looks back, and the scene in that direction is the same one from before. Earlier world models often failed this test. Look left, look right, look left again, and the left side had changed.

Two details matter here:

- **No game engine.** Some world models stay consistent by letting a game engine track every object. This one has none. The video model itself holds the memory.
- **Limited length.** A session can continue for about a minute, which is long for generated video but short for anything like a game level.

It also isn't instant. In the test, each step showed a "simulating what happens next" stage before the video picked up where it left off. That makes it more iterative than real-time systems like Google's Genie, which respond as you move.

## Can you run it on a single RTX 4090?

Yes, with a catch. Tencent says the model can run on one RTX 4090, a high-end consumer graphics card (GPU, the chip that does the heavy math for AI). But that is a compressed, slower version of the model. Compression usually means trading some speed and quality for a smaller memory footprint.

The setup Tencent used for real-time performance was much bigger.

| Setup | Hardware | What you get |
|---|---|---|
| Local, compressed | One RTX 4090 | Runs on a high-end gaming PC, but slower, and a compressed version of the model |
| Real-time | Eight data center GPUs | The experience Tencent tested for real-time interaction |

The sources don't give exact memory figures, file sizes or generation times for either setup, so check the project's own documentation before you plan a build or buy anything. If you don't own an RTX 4090 (a card that costs more than many laptops), the practical options are the project's hosted demo, if one is available, or renting cloud GPUs.

The "real-time needs eight data center GPUs" detail is the useful reality check. Interactive world generation is still expensive. A single consumer card gets you in the door, not the full experience.

## Is it really free to use commercially?

Per the source, the code and the model are both free to use, and Tencent allows commercial use. That's unusual for this kind of tool. Many impressive world models stay locked behind a company's servers or a research-only license.

Two cautions before you build a business on it:

1. **Read the license text yourself.** "Open" and "commercially usable" are summaries. The exact terms are in the license file in the project's repository.
2. **Treat the quality claims as unverified.** Tencent hasn't published numbers comparing the model with other world models. The consistency claims come from Tencent's own demos, which are naturally the best examples.

## How good is the quality?

Not good enough for final work, based on one creator's test.

He chose a third-person view, a lush green forest in the future, and a tiny ant as the subject. The model expanded his short prompts on its own, adding more detail to the subject and scene. The ant walking through the forest worked. Then he asked for a deer to run by, and the deer looked, in his words, a bit wonky.

Camera control was the weaker part. You can set the view with the arrow keys and WASD keys, but the mapping felt off: pressing W sent the camera right, and pressing A felt like moving forward. When he rotated the camera to see the deer from the front, the result was again wonky, though the camera angle did change as expected. His verdict: not the most impressive result, and not usable in final production.

That is a single test with short prompts, so it's a data point, not a benchmark. But it matches the general state of open video models: the idea works, and the polish isn't there yet.

## What is it actually useful for today?

The most believable use is **planning shots before you film them**. You talk your way through a scene, try different things (a deer, snow, a camera swing), and map out the sequence before anyone picks up a camera or opens an animation tool. Rough, fast, and cheap beats polished here.

Other realistic uses:

- Storyboards and pitch material where a rough moving visual helps
- Testing how a scene might feel from different camera positions
- Experimenting with interactive video as a technique, since the model is open

Not realistic yet: finished ads, film scenes, or anything where a strange-looking animal would hurt you.

## What does this mean for you?

**If you work in marketing, video or design:** this is a pre-visualization toy for now, not a production tool. Try the hosted demo if one is offered, keep your expectations modest, and watch how fast the quality improves. Open models tend to get better quickly once developers can modify them.

**If you're technical or curious about local AI:** the single RTX 4090 route is worth trying if you already own the card. Don't buy hardware for it yet. The compressed version is slower, and the real-time experience needs far more than a gaming PC.

**If you manage a team or budget:** the commercial license matters more than the demo. An open, commercially usable world model means a vendor can't be the only gatekeeper to this kind of tool. Have someone read the license before anyone builds a workflow on it.

**A practical habit for anyone testing these tools:** keep a log. Write down the prompt, the model, the hardware and what came out. Results from these models vary a lot, and notes are the only way to tell whether a new version is better.

## Could you build your own tracker for testing tools like this?

If your team is comparing AI video tools, a small internal app can replace the messy spreadsheet: a form to log each test (tool, hardware, prompt, result, a rating, a link to the clip), a gallery view, and a shortlist for shot planning. With [Remy](https://goremy.ai), you describe that app in plain English and it builds and deploys it with a database and sign-in, so your team can use it from a shared link. The code is standard TypeScript in a git repo you own. It is a small tool, but it keeps your testing notes in one place instead of scattered across chats.

## Frequently Asked Questions

### Can I run Tencent's world model on a single RTX 4090?

Tencent says yes, but it is a compressed, slower version of the model. The real-time setup it tested used eight data center GPUs. The sources don't specify memory requirements or generation times, so check the project's documentation for current numbers.

### Can I use it for commercial projects?

According to the source, the code and model are free to use, including commercially. Confirm the details in the license file in the official repository before you rely on it, since summaries can miss conditions.

### How long can the generated video run?

It builds video in 5-second segments and can continue for about a minute. Each segment takes your latest text or camera direction and looks back at what was already generated to keep the world consistent.

### Is it better than other world models?

That isn't established. Tencent hasn't published numbers comparing it with other world models, and the consistency claims come from its own demos. In one hands-on test, the output, including a deer and some camera moves, looked wonky.

### What's the difference between this and Google's Genie?

The source describes Tencent's model as more iterative: it simulates what happens next in steps, then continues the video. It is less real-time than Genie. Tencent's model is also open for anyone to download and use, which is its main advantage.
