Music 3.0 swaps the one-tag prompt for a timeline that keeps AI songs on track

An AI-generated song can sound finished and still betray the brief. A specified instrument fades by the second verse, the preset mood sags near the bridge, the vocal style never comes back. The track plays like a complete song, just not the one anyone asked for.

That failure mode is what Music 3.0, the new model in the Qwen music generation line, targets. Instead of one global tag covering the whole track, it uses a Structured Caption: a time-sequential description of what happens and when.

One tag, then the drift

Most text-to-music systems compress a request into a single global description of the track, and it holds only until the music starts to move: instruments drop out of the arrangement, the emotional tone flattens, and the vocal delivery survives only where the prompt happened to describe it. Music 3.0 replaces that single tag with a structured, time-sequential caption. It still covers genre, BPM, time signature, key, intended use, and production texture. It also tracks how emotion rises and falls, when lead and supporting instruments enter or exit, how the groove and the bass develop, and how singing style, harmony, and vocal effects change from section to section. Section labels in the lyrics, like [intro], [verse], [pre-chorus], [chorus], [bridge] and [outro], supply the macro structure; the caption fills in what changes inside each one.

The same complaint shaped the previous model in the line. Our coverage of Qwen-Music noted that end-to-end systems composing melody and arrangement as one task end up crowded and wandering. Qwen-Music planned the melody first. Music 3.0 extends that logic to the whole description layer.

Inside the Hybrid-LM: structure at 8B, detail at 0.6B

On the model side, Music 3.0 splits the work between two networks. An 8B Global LLM predicts, frame by frame, the tokens carrying the core semantic and structural content, holding whole-track context in mind. A 0.6B Local LLM works inside each frame, predicting acoustic tokens along the depth axis to add local sound detail. Most fine-grained detail doesn't need the biggest model; that is the same bet Nvidia makes with Nemotron 3.5 Lightning.

That division keeps time stable for songs up to five minutes while preserving variation inside sections. It extends the design from the Qwen-Music technical report, where a tokenizer compressed audio into a 25 Hz stream of semantic tokens and a melody chain-of-thought sketched a coarse vocal contour before full generation. The goal is the same: plan first, detail after.

The demos read like a proof of concept

The explainer ships with more than a dozen demo songs, each carrying the new caption format. A warm Mandarin pop ballad is described as 74 BPM in A-flat major, tender nostalgia opening into a celebratory chorus, a mature male baritone-tenor, delicate guzheng, soft strings, and a natural live-room sound. The arc, the instruments, the vocal qualities: stated up front, not left to inference.

The pattern holds across genres. "Cloud copy of me," a melodic EDM piece about uploading memory to a data stream, is captioned as reflective verses building into an uplifting hook-driven chorus, with glitch textures and a polished festival mix. The parent-duet Mandopop ballad goes furthest, specifying singers around 50 to 70, a strong Taiwanese accent, a mother who leads, a father who harmonizes, and an arrangement that builds from piano and strings to drums and electric guitar before fading to an acoustic close. None of this proves the model follows such captions faithfully; it shows what the format is designed to capture.

What section-level control changes for producers

For people who prompt music models, this shifts the point of control. Correcting a drift usually means regenerating the whole song or rewriting the global prompt and hoping. A caption that names what changes where, anchored by section labels, turns "what happens in the bridge" into a condition the model can be held to. The explainer is blunt about the goal: making where things change an explicit generation condition.

Competitors circle the same problem. MiniMax framed the Music 2.6 launch around four creator stories rather than a list of model gains, while its bigger architectural bet is H3, a single model for text, image, video, and audio.

Stability AI has locked in licenses with UMG and WMG and now has to ship on them. Music 3.0 aims at the production problem directly: generate the song that was described, for five minutes straight.

What the explainer doesn't say

No evaluation data shows how faithfully Music 3.0 follows a dense temporal caption; the demos illustrate intent, not adherence. Qwen-Music's report claimed state-of-the-art results against Suno V5 and MiniMax Music 2.6 in objective metrics and human preference, but caption-following is a different thing to score. Whether the model reliably honors long sequences of temporal instructions, and how caption intent interacts with lyric sentiment, stays open.

Open weights too. Qwen-Music's report committed to no release timeline, though the Qwen family's open-source history makes a release likely. The team has kept shipping open tooling in other domains, like the Qwen-MM-Plugins pack for coding agents.

The Music 3.0 material says nothing either way. And a release alone wouldn't settle much: Kimi K3 is the biggest open model ever, and it still isn't the best.

Where Music 3.0 lands, and how well it holds a five-minute brief, nobody has demonstrated yet.