HKGAI Introduces YuE2 | Open-Source Music Model WildSongBench Outperforms Suno

HKGAI, the Hong Kong Generative AI Research Centre, announced on September 18, 2026, the release of the open-source music large model YuE2. The team stated that it is the first open-source music large model globally to unify composition, cover versions, and editing through "symbolic planning"; it also ranked 4th on Hugging Face's full-site Trending list (Trending Score 595), the only music-related model in the top five. Within nine days of launch, the official main model recorded 11,626 downloads; including ComfyUI packages, community GGUF/quantized versions, and toolchains, the ecosystem saw a total of approximately 179,000 downloads.

HKGAI was established in October 2023, funded by the AIR@InnoHK platform focusing on artificial intelligence and robotics technology, and led by the Hong Kong University of Science and Technology. The research covers language, vision, and audio domains. YuE2 is regarded as a significant milestone in audio generation for the centre; the press release also likened the open-source model's momentum to the "DeepSeek moment" in the music industry, indicating its potential to match or even surpass some commercial closed-source systems.

WildSongBench: best-of-8 Comprehensive Scores Outperform Multiple Commercial Models

The evaluation benchmark WildSongBench conducted automatic assessments using 192 prompts (September 12, 2026). The table shows YuE2 (best-of-8) with a SongBench Avg of 6.9632, temporarily leading the ranking; a single YuE2 score was 6.7316. In comparison, Mureka 9 scored 6.9377, Suno v5 scored 6.8721, Suno v5.5 scored 6.7150, Suno v6 scored 6.5562, and the previous YuE 1 scored 4.9165. Other metrics were not necessarily top: for example, Suno v5 had the highest MuLan score (0.5428), LeVo 2 had the highest AudioBox PQ score (8.3966), and Suno v4.5 had the lowest PER score (5.80%).

System	SongBench Avg ↑	AudioBox PQ ↑	MuLan ↑	PER ↓
YuE2 (best-of-8)	6.9632	8.2714	0.5051	9.79%
Mureka 9	6.9377	8.0226	0.4394	11.69%
Suno v5	6.8721	8.1698	0.5428	8.10%
YuE2	6.7316	8.2598	0.5068	8.44%
Suno v6	6.5562	8.1296	0.4916	7.58%
YuE 1	4.9165	7.8683	0.2623	36.38%

White-box Control: Symbol Planning and Acoustic Rendering Decoupled

YuE2 separates "symbolic planning" and "acoustic rendering" using a mixture of experts (MoE) design to enhance controllability: AR experts use causal attention for autoregressive prediction of ABC scores and semantic skeletons; NAR experts use flow matching to predict acoustic latent representations, which are then decoded by a VAE to output 48 kHz stereo audio. Unlike the "prompt and generate" black-box sampling approach, creators can intervene at each step—for example, fine-tuning a single note without redoing the entire song, or changing the style while preserving the original melody.

The official demo uses the Mandarin pop song "Last Bus" for nine rounds of conversational editing, including reharmonization, arrangement, incorporating the melody of "Twinkle Twinkle Little Star," developing variations, and rewriting lyrics, resulting in a complete English jazz arrangement. Each step corresponds to verifiable score and audio, rather than an unrelated random generation.

Six Languages, Sixty Genres, Local Open-Source Ecosystem

In terms of capabilities, YuE2 supports six languages including Chinese, English, Japanese, Korean, Russian, and Spanish, covering over 60 genres (including City Pop, heavy metal, jazz, and flamenco). The weights are open-sourced and freely downloadable (for non-commercial use only), and can be run locally on consumer-grade single GPUs; commercial licenses require separate arrangements. The team has also open-sourced transposition tools, music understanding models, and evaluation benchmarks, connecting the "transposition—creation—editing—evaluation" toolchain. Project entry: GitHub multimodal-art-projection/YuE.

• Open-sourced weights (non-commercial) + runnable on local single GPU
• Creation / cover / editing unified in symbolic planning workflow
• Open-sourced transposition, understanding models, and evaluation benchmarks simultaneously