AIBRIEF
A real time AI world model that generates interactive scenes on the fly outputs 24 FPS video and 48 kHz stereo in one pass, with model weights due on GitHub in the coming weeks.
Noiz AI released HelixWorld 1.0, a real-time interactive world model that generates 24 frames-per-second video and 48 kHz dual-channel audio together from a single shared Transformer state. A user walking, turning, or reaching into a generated scene updates the visual frame and the sound field at the same moment.
The team pairs Noiz AI engineers with researchers from Hong Kong University of Science and Technology, Tsinghua, Carnegie Mellon, and Google DeepMind, building on the LTX2.3 audio-video base with action-control fine-tuning. The training pipeline is the load-bearing piece: it fuses real first-person video, which carries spatial and material fidelity, with game-engine captures that supply precise camera-pose and spatial-audio labels. After a custom dual-channel, semantic, and spatial-alignment cleaning pass, Noiz AI reports millions of frame-aligned training segments.
The model enters a competitive field whose named rivals (Google DeepMind's Genie 3, Tencent's WorldPlay, and Ant's LingBot-World) still ship silent visuals. The launch demo places a viewer on a generated street, where footsteps, traffic, and a passing engine all react to head movement.
What remains unverified: independent benchmarks against those rivals, third-party audio fidelity tests, and the open-source drop itself. Noiz AI plans to release the model weights and code on the NoizAI/HelixWorld GitHub repo in the coming weeks, building on the prior AudioX-Turbo release with the same collaborators. Funding is single-source-reported: a seed round of several tens of millions of yuan (roughly $4–12 million at mid-2026 exchange rates) from GSR Ventures, per pedaily.cn. The launch is the announcement. The open release is the test.