Tencent drops Hy-Nomas, an interactive world model that runs on a 4090

Hy-Nomas is an open-source interactive video world model from Tencent Hunyuan that lets you generate and explore virtual environments in real time using text prompts and camera trajectories.

Instead of rendering a fixed, non-interactive video clip, Hy-Nomas responds to multi-turn natural language commands. You can switch between first-person and third-person perspectives, steer the camera, spawn new objects, and alter the environment on the fly.

What it can do

Traditional video models generate a few seconds of video all at once. Hy-Nomas works differently: it generates short video chunks autoregressively while keeping a persistent KV cache across segments.

This lifelong memory allows it to remember past events and maintain scene consistency as you roam around. In tests, the world remains spatially and visually stable even across continuous runs lasting over a minute.

The model outputs video at 720p resolution and 24 frames per second. You can generate scenes purely from text or upload a reference image as an initial state and then continue exploring via keyboard navigation and text prompts.

Beyond entertainment and interactive roleplaying, Tencent notes that the model's visual consistency makes it suitable for generating simulation environments for robotics training.

How to run it

Getting a real-time world generator running locally usually requires a data center cluster, but Hy-Nomas is surprisingly accessible.

Tencent released two main variants. The base model weighs around 34GB, while a distilled INT8 checkpoint trims that down to roughly 9GB. The distilled version cuts inference down to just 4 to 8 denoising steps.

Thanks to these low-memory inference techniques, the 9GB version runs on a single consumer RTX 4090 GPU. For faster multi-chip setups, community benchmarks showed real-time 720p generation on Google TPU v6e-4 slices without dropping frames over 60 seconds.

The project is fully open under an Apache 2.0 license. The GitHub repo includes inference code, training scripts, LoRA fine-tuning workflows, and data annotation tools. Third-party integrations like ComfyUI, SGLang, WanGP, and Reactor Runtime have already added WASD keyboard camera controls.

You can test it out right away via the interactive web demo on GitHub or pull the weights to deploy it locally.

Why it matters

World models like Google's Genie have captured the imagination of game designers and robotics engineers, but they have largely remained locked behind closed doors and massive server farms.

Hy-Nomas shifts that dynamic by proving that multi-turn, interactive 720p world generation can fit onto a consumer-grade 24GB card. Giving creators and researchers full Apache 2.0 access to weights and training scripts will dramatically lower the barrier for interactive media and robot simulation experiments.