A songwriter hums a hook, types a title and a few lines, picks a reference track, and waits. Sixty seconds later there is a finished 80-second stereo song — vocals, instruments, and mix — in a voice that sounds native, not translated.
That is the promise of SkyMusic (天工SkyMusic), the AI music model from Kunlun Tech (昆仑万维, 300418.SZ), and it arrived with a claim its maker backed with a score.
The day it opened
On 17 April 2024, to mark the first anniversary of its "Tiangong" (天工) base model, Kunlun Tech opened public testing of SkyMusic (天工SkyMusic), its music large model (音乐大模型), alongside Tiangong 3.0. Kunlun called it China's first music SOTA model — and said it was the first time a self-developed Chinese model led the world in an AIGC vertical.
The base model behind it, Tiangong 3.0, carries 400 billion parameters and, the company said, surpassed Grok-1's 314 billion to become the largest open-source MoE (Mixture-of-Experts) model at the time.
The Suno V3 scorecard
Kunlun put SkyMusic head-to-head with Suno V3, then the best-known AI music model. In its cited evaluation, SkyMusic scored 6.65 overall and led in vocal-and-BGM quality, vocal naturalness, and pronunciation intelligibility.
The reason matters more than the number. Most earlier AI music research chased symbolic generation — MIDI-like note sequences — which top out at faceless background music. SkyMusic took a different route: a Sora-like architecture combining a large-scale Transformer with a Diffusion Transformer to generate instruments, vocals, melody, volume, and notes end-to-end. That is technically harder and is why Kunlun says only a handful of players worldwide attempt it.
What it actually does
SkyMusic's five headline abilities:
- High-quality generation — up to 80 seconds of two-channel stereo AI song.
- Vocal synthesis — the company says its Chinese singing clearly beats foreign products, where tonal languages usually break.
- Lyric control — precise section-by-section structure so emotion lands where the writer intends.
- Style range — rap, folk, electronic, and more; upload a reference track and it matches the style, or let it pick.
- Dialect support — Cantonese, Chengdu, Beijing and other dialects, a feature that doubles as cultural preservation.
The user flow is deliberately flat: name plus lyrics (the AI can draft lyrics if you are stuck), a reference track or auto-style match, then automatic composition and mixing. Kunlun had already put AI-composed music on Spotify, QQ Music, and NetEase back in 2022, so SkyMusic is an acceleration of a long-running lab effort, not a pivot.
Why "sounds Chinese" is the point
English-first models mangle Mandarin tones — the difference between "mother" and "horse" collapses into noise. SkyMusic's vocal edge on Chinese is the feature that separates it from Suno for local creators. Dialect support pushes further: a Guangzhou label can ship a Cantonese track without booking a Cantonese vocalist.
This is also why Kunlun frames SkyMusic as the moment Chinese self-developed models led globally in a creative vertical, not just matched. The claim is about a category (AI music) where, until then, no product had triggered a "music ChatGPT moment."
The open question
SkyMusic is "publicly usable," but usable is not the same as studio-ready. An 80-second stereo clip is a strong demo; a three-minute song with consistent character across a full album is a different bar. And the field is moving fast — foreign and domestic rivals are shipping music models on short cycles, so a 6.65 lead in April 2024 is a snapshot, not a fortress.
Why this matters beyond China
Music is the next frontier for generative media after text and images. When a publicly listed Chinese internet company open-tests a music model, it signals that AI composition is moving from novelty to a feature inside existing music and social products. The open question is rights: who owns a song a model helped write, and how platforms label it.
Honest limitations
The 400-billion-parameter claim, the Suno V3 comparison, and the 6.65 score all come from Kunlun Tech's own launch materials and securities-research notes citing them; there is no independent blind test here, and vendor-led head-to-heads carry built-in bias. "Largest open-source MoE" was true at launch and may since have been passed by newer models. The 80-second limit is a hard ceiling on a single generation, not evidence of full-song capability. Dialect quality will vary by dialect and by how much training data each one received. And like all AI music, rights and likeness questions around voice and reference tracks remain unresolved.
What readers can do now
- If you make content, prompt SkyMusic (天工SkyMusic) with a real brief — a 30-second jingle in your brand voice — and judge whether the vocal passes as native Chinese.
- If you work in a dialect market, test a Cantonese or Sichuan-track and check intelligibility against a human singer.
- If you track the field, re-run the Suno comparison yourself each quarter; a single 6.65 score from 2024 tells you where SkyMusic started, not where it stands.