Tencent Open-Sources Hy-Nomas: Interactive Video Generation with WASD Controls on a Single RTX 4090
Most video models produce a clip in one go, so turning the camera on the tenth second means starting the generation over. Tencent has open-sourced Hy-Nomas, which generates video in segments and lets you move the camera, add objects, and set events with text between them. The model keeps track of earlier frames, so the world isn't redrawn from scratch at every step. A quantized configuration exists for a single gaming RTX 4090, but no speed figures have been published.
AI-processed from TG: @neurogen_news (Neurogen); edited by Hamidun News
Most video models work like a typewriter without a backspace key: you write a prompt, wait, and get a finished clip. If you want the camera to turn right on the tenth second or a car to appear in the frame, your only option is to change the prompt and run everything again. Tencent has released Hy-Nomas as open source, and it works differently: it generates video in parts, and the user controls what comes next.
You can move through such a world with the WASD and arrow keys, as in a regular computer game (W, A, S, D mean forward, left, back, and right). According to the authors, the model can run on a single RTX 4090, a top gaming graphics card with 24 GB of memory that sits in home PCs rather than data centers.
How it works in practice
Imagine a game scene: a hero stands in a square. You generate the first fragment, look at it, and decide what to do next. You can shift the camera, add a fountain to the square, change daylight to evening, or type that a cart is rolling out from around the corner. The next segment is built with your instructions already taken into account, and you don't need to restart generation from scratch.
That makes Hy-Nomas closer to an interactive "world model" than to a classic video generator. You don't order a finished video; you steer it as it unfolds, and in the available integrations you do this with the same WASD and arrow keys.
The main technical problem with such systems is familiar to anyone who has watched long AI videos: after a few seconds a building changes shape, a road changes color, and objects disappear. According to the developers, Hy-Nomas uses the generation history, meaning it relies on fragments it has already created. As a result, the world stays coherent and isn't redrawn at every step. How well this holds up in long sessions can't be judged from the description alone; the demo and the repository are there for testing.
Running on a single graphics card
For enthusiasts, the most interesting detail is the configuration for a single RTX 4090. It is achieved through quantization, which means storing the model's weights in a more compact format. Some of the computational components are moved into the computer's system RAM. This is a familiar technique: it makes it possible to fit a large model into 24 GB of video memory, but it usually costs speed, because data has to shuttle between system memory and the graphics card.
Caution is warranted here. The developers do not state a specific speed for the 4090. Moreover, frames per second (FPS) and video generation speed are different measures. A model may show smooth motion in a finished clip while taking minutes to create that clip. For interactive control, what matters is response time: how long passes between pressing a key and the appearance of the next fragment. Such figures are not available yet, so calling the 4090 setup "playable" would be premature.
What developers and product managers should do
First, try the demo to see where the line between video generation and interactive simulation falls today. For prototypes of game scenes, film and advertising previsualization, training simulations, and level concepts, the idea of step-by-step control looks more practical than endless prompt tuning.
Second, the open code lets you test the claims on your own hardware. If you have an RTX 4090, measure two things: how many seconds pass from a command to the next fragment, and how long the world stays coherent. These numbers determine whether the model suits a product, not the impressiveness of the demo.
Third, it's worth watching how Hy-Nomas is compared with other world models. The approach in which you steer generation instead of ordering it whole is gradually becoming a distinct class of tools. Tencent's open implementation lowers the barrier to entry: researchers and small studios get a base for experiments without waiting for closed services.
For now, the picture is this: the idea is clear, the code is published, and running on consumer hardware is claimed, but the authors have not reported the key parameter, speed on the 4090. For a fresh release that is normal, but the model should be judged by your own measurements rather than by the description.
- Nov 21, 2026Krea Opens Beta of Krea Bots: 9 AI Bots Produce a 15-Second Commercial
- Nov 21, 2026Mozilla's GPT-6 Cyber Agents Found and Fixed 8x More Firefox Bugs in a Year
- Nov 21, 2026Meta* Releases XRStereo: Nearly 18,000 Stereo Frame Sets for 3D Vision
- Nov 21, 2026Alibaba's Wan 3.0 Tops the Video Leaderboard at 1175 Elo, $0.20 per Second
Key Questions Answered
How it works in practice?
Imagine a game scene: a hero stands in a square. You generate the first fragment, look at it, and decide what to do next. You can shift the camera, add a fountain to the square, change daylight to evening, or type that a cart is rolling out from around the corner. The next segment is built with your instructions already taken into account, and you don't need to restart generation from scratch. That makes Hy-Nomas closer to an interactive "world model" than to a classic video generator.
What are the key details on Running on a single graphics card?
For enthusiasts, the most interesting detail is the configuration for a single RTX 4090. It is achieved through quantization, which means storing the model's weights in a more compact format. Some of the computational components are moved into the computer's system RAM. This is a familiar technique: it makes it possible to fit a large model into 24 GB of video memory, but it usually costs speed, because data has to shuttle between system memory and the graphics card. Caution is warranted here.
What developers and product managers should do?
First, try the demo to see where the line between video generation and interactive simulation falls today. For prototypes of game scenes, film and advertising previsualization, training simulations, and level concepts, the idea of step-by-step control looks more practical than endless prompt tuning. Second, the open code lets you test the claims on your own hardware. If you have an RTX 4090, measure two things: how many seconds pass from a command to the next fragment, and how long the world stays coherent.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.