MiniMax releases Music 3 for five-minute songs with hierarchical generation
Yan Junjie's model splits long-range composition from acoustic rendering and accepts detailed sectional controls, though its official weights remained unavailable at verification.
By RuntimeWire Staff · Published
Primary source: MiniMax
Why it matters
Music 3 gives developers a detailed view of a five-minute song model that separates long-range composition from acoustic rendering. Its repository and demo expose the architecture, input controls and selected outputs, while the official weights were not yet available at verification.
MiniMax, the Shanghai-based AI company founded by Yan Junjie, announced Music 3 on August 13, 2026, describing it as an open-weights model for complete songs of up to five minutes from lyrics and a detailed production brief. A public demo and technical materials were available at launch. The official repository contained documentation and assets, while the official release materials did not yet make the model weights available. MiniMax also published a ModelScope listing.
The announcement pointed developers to a Hugging Face model page, a GitHub repository, ModelScope and a public demo. MiniMax calls Music 3 "production-ready," a characterization supported for now by company-selected samples and a reproducible generation example rather than independent quality tests, disclosed customers or professional-studio adoption.
The release follows the technical path Yan took after more than six years at SenseTime, where he rose to vice president and deputy head of its research institute. MiniMax's management biography says Yan earned a doctorate in artificial intelligence from the Institute of Automation, Chinese Academy of Sciences, and later conducted postdoctoral research at Tsinghua University. MiniMax's annual report identifies him as chairman, CEO and CTO, placing model research and corporate strategy under the founder.
A song model built around hierarchy
According to the official model page, Music 3 separates composition from acoustic rendering. An 8-billion-parameter Global LLM, initialized from Qwen3-8B, handles long-range structure and predicts the first semantic music codebook. A 0.6-billion-parameter Local LLM fills in frame-level acoustic detail across seven additional codebooks.
The hidden states from those models feed into a 2.4-billion-parameter Flow Matching module and a 123-million-parameter Flow-VAE decoder. Music 3 produces 32 kHz, 16-bit stereo WAV files. MiniMax says the architecture helps maintain rhythm, vocal identity, musical themes and arrangement changes across a full song.
Users supply two inputs: lyrics and a music description. Lyrics can carry section labels including [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo] and [Outro]. The description can specify genre, BPM, key, emotional progression, vocal timbre, backing vocals, instruments, percussion and changes in arrangement.
That interface lets a developer request a restrained first verse, introduce drums in the chorus and reserve a guitar solo for the bridge. The controls give Music 3 a song-production vocabulary that extends beyond a single text prompt and lets users describe changes across a five-minute arrangement.
The tokenizer described in the model materials uses eight residual-vector-quantization codebooks. One 16,384-entry codebook represents semantic content, while seven 1,024-entry codebooks carry acoustic detail. The inference instructions describe generation at 25 audio frames per second.
What developers can test
MiniMax described Music 3 as an open-weights release, but the distribution was incomplete when checked on August 13. The official repository contained documentation and demo assets, while the official weights were not yet available.
The demo establishes the input format and provides selected outputs.
The available model materials also describe the intended inference setup. The model card says CPU offloading can reduce video-memory usage to roughly 8 GB. Repository instructions require CUDA, describe non-streaming generation and call for two GPUs.
The company behind Music 3
MiniMax was founded in early 2022 and develops models and products spanning text, audio, images, video and music. MiniMax says its models and products have served more than 300 million individuals across over 200 countries and regions, along with more than 1 million enterprises and developers across over 100 countries and regions. Those are company-wide figures; MiniMax has not broken out Music 3 adoption.
Alibaba led a $600 million financing round in March 2024 at a valuation above $2.5 billion, with HongShan also participating, according to Fortune. MiniMax subsequently listed in Hong Kong in January 2026 and raised approximately $619 million in its IPO, according to Reuters.
MiniMax's 2025 annual report recorded $79 million in revenue, up from $30.5 million in 2024, along with $252.8 million in research and development expenses. MiniMax's annual report listed 418 employees at the end of 2025, including 415 full-time and three part-time workers. Those figures cover MiniMax Group rather than the Music 3 team.
Music 3's five-minute output puts it in range of another downloadable music-model family. Stability AI's Stable Audio 3.0 is an open-weights comparison that supports compositions of roughly six minutes, according to the company's materials.
Hosted competitors such as Suno and Udio spare users from managing local inference infrastructure. Music 3's model materials describe an intended local inference setup, though its official weights were unavailable when checked.
MiniMax is defending a US copyright lawsuit filed by Disney, Universal and Warner Bros. Discovery over alleged unauthorized training and outputs from Hailuo AI, MiniMax's image and video service. The case is not specifically about Music 3.
Yan is using Music 3 to place MiniMax's audio research in front of developers who want longer songs and detailed arrangement controls. Its hierarchical architecture and five-minute demos give them an initial basis for evaluation. The official weights were not yet available when the release materials were checked.