Skip to content
Neural Daily

ModelsNews

Tencent Hy-Nomas: an open interactive video world model

🇭🇺 Olvasd el magyarul →

Tencent Hy released Hy-Nomas, an interactive video world model, under Apache-2.0. You steer it segment by segment with text instructions and camera moves, and it runs on a single RTX 4090.

By: Mikola Máté Patrik · · 2 min read

Hy-Nomas title card: free-form interactive world generation, Tencent Hy logo over a blurred harbor scene with a hot-air balloon
Image: Tencent Hy · Press kit · Official project page og:image · Source

What Tencent released

Tencent's Hy team (formerly Hunyuan) published Hy-Nomas on Hugging Face on October 24 (Hugging Face). It is a video world model: instead of one clip, it builds a world in five-second segments, and the user can give a new instruction before each one.

Along the way you can introduce objects, switch style or scene, describe a character's next action or supply a camera trajectory. A persistent KV cache keeps earlier segments in context, which keeps the world consistent.

How it works

  • A segment-autoregressive diffusion transformer with 54 dual-stream MMDiT blocks.
  • Text is encoded by the language part of Qwen2.5-VL-7B, while camera trajectories enter through separate cross-attention.
  • A distilled model: the Diffusers example uses 4 denoising steps per segment at 24 fps, and the project page shows long-form 720p video.
  • With an INT8 DiT and FP8 decoder it runs on a single RTX 4090 (
  • [object Object]
  • ).
Hy-Nomas architecture diagram: structured prompt, video clip and camera trajectory feed dual-stream MMDiT blocks with segment-wise causal attention and a KV cache
Tencent Hy · Screenshot · Architecture figure from the official model card · Source

Image: the Hy-Nomas architecture. Source: Tencent Hy.

Tencent also released Hy-Nomas-Bench, with 280 test cases across four areas: multi-turn text interaction, camera navigation, interleaved text and camera control, and spatial memory. The repo includes training and inference code, LoRA fine-tuning and a web demo. The technical report is still marked as coming soon.

Why it matters

Interactive world models have mostly been closed demos, such as Google's Genie line. Hy-Nomas ships full weights under a license that allows commercial use, and it runs on a consumer GPU. Anyone working on game prototyping, simulation or synthetic data for robotics can now test what an open world model can do today. Without independent evaluations it is too early to judge quality; comparisons should follow in the coming weeks.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles