An open interactive audio-visual world model from Noiz AI and HKUST: real-time 768×512 video at 24 fps with spatial stereo sound, steered with keyboard controls, running on a single H800. Released as a preview on 12 October 2026.

Can I run Helix World myself?

The weights are on Hugging Face as NoizAI/HelixWorld-preview and the demo is at helixworld.org. Real time needs a data-centre GPU; the paper quotes an H800.

Helix World: real-time interactive AI video world

A world model is a video model you steer: instead of a prompt and a clip, a scene that keeps rendering as you move through it. Noiz AI and the Hong Kong University of Science and Technology released Helix World preview on 12 October 2026, and it is the first open one with sound.

What it does

Generates 768×512 video at 24 frames a second in real time while you control the camera with WASD or the arrow keys. Audio is generated with the picture: stereo spatial sound at 48 kHz, so a source to your left is heard on the left and moves as you turn.

The numbers

A real-time factor of 0.77 on a single H800: the model generates a second of video in 0.77 seconds, with a margin for the interaction loop. It works in 1.3-second chunks, each conditioned on the last and on your input.

What it is built on

LTX-2.3, Lightricks' open video model, which is why the frame rate and the resolution look like LTX's. The weights are on Hugging Face as NoizAI/HelixWorld-preview, the paper is arXiv 2609.38123 and there is a browser demo at helixworld.org.

Where it sits

Google's Genie and the game-engine world models from the large labs are closed. Helix is the first with open weights that does audio, and the first you can run on one card.

The sound is the part to pay attention to. Every open world model until now has been silent, and silence is why they read as screensavers rather than places; spatial audio that turns with the camera does more for the sense of being somewhere than another hundred pixels of width would. The 0.77 real-time factor is the number to be sceptical about, because it is one H800 for one user, and the chunked architecture means a 1.3-second latency to any input you give, which is fine for walking and hopeless for anything that needs reflexes. It is a preview of a medium rather than a product, and a good one.