HelixWorld
Released Sep 3, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Type
- World Model
- Modality
- Multimodal
- Context window
- —
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Sound is generated with the picture rather than dubbed on afterwards — a drummer who pauses on screen goes quiet instead of playing on, which post-hoc video-to-audio pipelines get wrong
- The stereo field turns with the camera, so a source you pan past moves from one ear to the other instead of staying fixed in the mix
- Steered like a game from a single starting image and prompt — WASD and arrow-key actions, combinable, rather than re-prompting for every shot
Weaknesses
- Running the open checkpoint takes a single 80 GB-class datacentre GPU, so it is out of reach of consumer cards
- Controls land a beat late — input binds to the next generated chunk of world, so turns and prompt changes arrive a moment after you press them
- A preview checkpoint with no independent hands-on testing yet — every quality claim, including the blind user study, is Noiz's own, and Noiz lists long-horizon consistency as future work
Not an LLM: an interactive audio-visual world model that takes a first-frame image, a text prompt and a stream of camera/keyboard actions and generates 768x512 video at 24 FPS with 48 kHz camera-aligned stereo sound, so none of this site's benchmarks apply and every cell is null by design; context, output and pricing are likewise not applicable. Built by Noiz AI with HKUST on the open LTX-2.3 backbone (Gemma 3 12B text encoder), as a bidirectional teacher distilled into a few-step streaming student; Noiz reports a steady-state real-time factor of 0.77 on one H800 including decoding, about 70 GB VRAM in BF16, and has published no parameter count. Announced as HelixWorld 1.0 on 2026-08-17 (QbitAI) with weights promised 'in the coming weeks' and the API in internal testing; releaseDate is 2026-09-03, when Preview v1 inference code (Apache 2.0, GitHub NoizAI/HelixWorld) and checkpoint (Hugging Face NoizAI/HelixWorld-preview) shipped alongside a browser demo at helixworld.org. The checkpoint's LICENSE is Apache 2.0, though the card also calls it a 'private inference checkpoint' and says repository access 'does not replace acceptance of the applicable model license'; the technical report (arXiv 2609.38123, 2026-09-29) is CC BY-NC-SA 4.0, which covers the paper only. availability is self-host because there is no public API, no Hugging Face inference provider, and the helixworld.org demo states no access terms. Noiz's HelixBench (released 2026-10-12) and WBench comparisons are self-reported and given only as figures. Company details conflict: Pedaily names the registered entity as Shenzhen-based and founded in 2025, while Dealroom gives Hong Kong, 2024; the registered-entity figures are used.
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor