CONFIRMED (release) / SELF-REPORTED (results): Helix World is out, with a paper, code, a preview checkpoint and a public demo. Every benchmark figure below comes from the authors, including scores on HelixBench, a benchmark they built themselves. TSN has not tested the model.
Noiz AI and the Hong Kong University of Science and Technology (HKUST) have released Helix World, an interactive “world model” that generates video and matching stereo sound at the same time, in real time, as a user moves through a scene [1][2]. “When the camera turns, the sound field turns with it,” Noiz AI wrote in its announcement on 12 October [1]. (CONFIRMED)
What a world model is
A world model, in this sense, is an AI system that generates an explorable scene on the fly: instead of playing back a fixed video, it produces the next moments of picture in response to the user’s controls. The authors say most such models are silent [1][2]. Helix World is their attempt to add sound that matches what is on screen and where it is.
How it works
You start from a single image, move with the WASD keys and look around with the arrow keys; a text prompt is optional [1]. The model generates about 1.3 seconds at a time, and a new key press or prompt takes effect from the next chunk [1]. The authors say it runs at 24 frames per second at 768×512 on a single NVIDIA H800 GPU, with a real-time factor of 0.77 including video and audio decoding, meaning it needs 0.77 seconds of compute per second of output [1][2]. (SELF-REPORTED)
It is built on the open-source LTX-2.3 audio-visual model [1][2]. The training data, according to the authors, started as 4,100 hours of raw footage (real-world video, game recordings and open-source datasets) and was filtered to 3,000 hours in 2.1 million clips with true stereo sound and camera positions [1][2].
The paper, “HelixWorld: A Real-time Interactive Audio-Visual World Model” (arXiv:2609.38123, dated 29 September 2026), is on arXiv [2]. Code is on GitHub (NoizAI/HelixWorld), a preview checkpoint is on Hugging Face (NoizAI/HelixWorld-preview), and the demo is at helixworld.org [1][3][4][5]. (CONFIRMED: all four were live when TSN checked on 12 October)
The results, and where it is behind
All of these figures are the authors’ own; none has been independently verified.
- Picture quality (WBench navigation): Helix World scores 79.9. That is behind Alaya-EVOKE-Turbo at 82.0 and EchoWM at 81.0 (Zing-0.5 also scores 81.0), and ahead of SANA-WM at 76.0 and Matrix-Game 3 at 71.3 [1][2]. The authors’ point is speed: the three higher-scoring models need 1.98 to 4.29 seconds of compute per second of video on the same GPU, by their measurement [1][2].
- Sound that follows the camera (HelixBench): Helix World scores 41.8 on “spatial-acoustic consistency”, a measure of whether left-right stereo tracks where sound sources are on screen, against 12.6 for EchoWM, the only other world model in the comparison that makes sound [2]. HelixBench, 1,015 human-verified test cases across six scene types, was created by the same team and released with the model [1][2]. The blog’s text gives EchoWM’s score as 12.7; the paper’s Table 4 says 12.6 (12.6476) [1][2].
- Not best everywhere on sound: the authors note that other systems had lower audio-video timing offset and one had slightly higher text-audio alignment [1].
- User study: with sound on, participants preferred Helix World over EchoWM 67.9% to 20.8%. With the sound muted, they preferred EchoWM, 44.7% to 26.3% [1]. “Its advantage comes mainly from the audio,” the authors wrote [1].
Limitations the authors list
- Output is 768×512 only [1].
- Real-time speed is reported only on an NVIDIA H800 data-centre GPU; no consumer GPU figures are given [1].
- The public checkpoint is a preview release [1].
- On visual quality alone it is slightly behind the best silent models (79.9 vs 82.0), and muted, viewers preferred EchoWM [1].
Who built it
The paper lists Lei Ke, Jiahao Pan and Zeyue Tian as equal-contribution authors, with Zeyue Tian and Wei Xue as corresponding authors, from HKUST and Noiz AI [1][2].
What is not known
- How it performs on benchmarks the authors did not design or choose.
- How fast it runs on consumer hardware.
- When a non-preview checkpoint will follow.
Sources
- Noiz AI, “Meet Helix World: A Real-Time Audio-Visual World Model”, 12 October 2026. https://noiz.ai/lp/blog/meet-helix-world
- Lei Ke, Jiahao Pan, Zeyue Tian et al., “HelixWorld: A Real-time Interactive Audio-Visual World Model”, arXiv:2609.38123, 29 September 2026. https://arxiv.org/abs/2609.38123
- Helix World demo. https://helixworld.org
- NoizAI/HelixWorld, GitHub. https://github.com/NoizAI/HelixWorld
- NoizAI/HelixWorld-preview, Hugging Face. https://huggingface.co/NoizAI/HelixWorld-preview
