From driving to flying: Tencent open-sources an interactive world model, and the Demo has no waitlist

Tencent Hunyuan has released Hy-Nomas, an interactive video world model. It accepts multi-turn text instructions and continuous camera control, generating variable-length video clips. The official release includes model weights, code, Bench, and Demo together, and the Demo does not require queuing.

The core message from the paper abstract is straightforward: an 8B diffusion backbone supports free-form text instructions and continuous camera control, generating across multiple interaction rounds. The model caches history, and the current input together with past clips determines the next video segment. Its design aligns structured video captions with action-aware causal attention: descriptive context can be accessed across clips, while action instructions are restricted to their expected time windows.

In everyday terms, the model remembers the scene, but a command like 'turn left' only takes effect in the segment where it should. In a reply, Saksham Jain noted that restricting each action prompt to its own time window is a good design; otherwise, a single 'turn left' would bleed into every later clip. This detail is critical for multi-turn interaction.

Another component is the VLM Prompt Enhancer, which translates user requests and multimodal interaction history into structured conditions for the next segment. Training uses a multi-stage curriculum that progressively introduces causal generation, text interaction, camera control, and model-generated history training, followed by diffusion distillation. The human evaluation in the abstract claims that Hy-Nomas outperforms existing world models.

In a reply, Rui Zhao shared a demo: driving, unfolding wings, taking off, landing, dodging tank fire, and escaping. Keyboard control and text prompts can be switched.

Replying to AK, author Weizhi Wang said that Hy-Nomas supports functionality similar to Genie4, with the difference that the model is fully open source. The open-source list includes model weights, code, Bench, and a lightweight adaptation pipeline, aimed at reproducible research on interactive world generation.

Links are here:

• Paper page: https://huggingface.co/papers/2610.36729
• Technical report: https://github.com/Tencent-Hunyuan/Hy-Nomas/blob/main/Hy_Nomas_Tech_Report.pdf
• Project page: https://hy-nomas.hy.tencent.com/
• Demo: https://hy-nomas.hy.tencent.com/demo
• GitHub: https://github.com/Tencent-Hunyuan/Hy-Nomas
• Model weights: https://huggingface.co/tencent/Hy-Nomas
• Bench: https://huggingface.co/datasets/tencent/Hy-Nomas-Bench
• Hugging Face Space: https://huggingface.co/spaces/tencent/hy-nomas

Rui Zhao also posted a simple demo of Hy-Nomas running in the Space. If you want to see the results first, you don’t have to wait for a list.

Open-sourcing to this degree is more meaningful than watching a single show-off video. Whether a world model can be reproduced and modified, and whether the weights and training pipeline are released, is where the real test will come.