Tencent's Hunyuan team has released Hy-Nomas, an open-source video world model that supports continuous interaction through free-form text prompts and camera trajectory control. Unlike traditional video generators that produce a fixed clip, Hy-Nomas builds the world segment by segment, maintaining consistency thanks to a key-value (KV) cache that carries previous context.
The launch took place around October 23, 2026, with a GitHub repository, weights on Hugging Face under the Apache 2.0 license, an official website, and an online interactive demo. The reference configuration generates 720p output at 24 FPS, with up to about 60 seconds of continuous duration and eight denoising steps per segment in the distilled version.
What the model enables
Users can:
- Direct actions, transform scenes, and change styles with natural language across successive turns;
- Control the camera alongside text, interleaving the two input forms;
- Maintain spatial and object consistency when returning to previously generated locations, thanks to world memory;
- Adapt the model with LoRAs, including one for robotic arm simulation trained on real-world data.
There is also Hy-Nomas-Bench, a set of 280 test cases that evaluates multi-turn interaction, camera navigation, interleaved control, and spatial memory. The full technical report has not yet been published.
Runs on consumer hardware
The standard configuration uses sequence parallelism (typically multiple GPUs), but the project includes an optimized path for a single NVIDIA GeForce RTX 4090. It uses INT8 DiT weights, an INT8 KV cache, an FP8 PixelUMM decoder, and CPU offload. This requires sufficient system RAM and adds transfer overhead, but makes local experimentation viable. The community has already begun porting the model to ComfyUI, with reports of use on GPUs with around 12 GB of VRAM in experimental setups.
The backbone is a dense multimodal diffusion transformer (MMDiT) with 54 dual-stream blocks, using a text encoder based on Qwen2.5-VL-7B and camera conditioning via cross-attention. Generation is autoregressive by segments, with causal attention between them.
Why it matters for games and AI
Interactive world models like Hy-Nomas point toward environment prototyping tools, dynamic content generation, and even simulations for robotics or agents. Instead of isolated clips, the focus is on persistent worlds that the user shapes as they are generated. This aligns with other fronts of generative AI for games, such as real-time rendering and agents that explore or build environments.
It is still early to talk about real-time playable generation on consumer hardware at the level of a traditional game engine — inference times depend on the configuration and resolution — but the release of the weights and code accelerates community experimentation.
Sources
- Official GitHub repository
- Weights and documentation on Hugging Face
- Official site and interactive demo
- Original announcement on X by Tencent Hunyuan
Transparency: This content was created, edited, or reviewed with the aid of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
