MiniMax ships Music 3 documentation and calls it production-ready. The weights are the missing part.

MiniMax announced Music 3 on August 13, 2026, describing it as an open-weights model that produces complete songs of up to five minutes from lyrics and a detailed production brief ^[1]. When the distribution was checked on the announcement date, the official repository held documentation and demo assets while the weights were not yet available ^[4], which makes this a roadmap signal rather than something you can put in a pipeline.

The architecture is the interesting part, and it is fully documented even without a checkpoint. An 8-billion-parameter Global LLM initialized from Qwen3-8B handles long-range structure and predicts the first semantic music codebook ^[6]. A 0.6-billion-parameter Local LLM fills in frame-level acoustic detail across seven more codebooks ^[7]. Hidden states from both feed a 2.4-billion-parameter Flow Matching module and a 123-million-parameter Flow-VAE decoder ^[8], for roughly 11.1 billion parameters across the stack ^[1]. Output is 32 kHz, 16-bit stereo WAV ^[9]. MiniMax says the split between composition and acoustic rendering is what holds rhythm, vocal identity, themes and arrangement changes together across a full song ^[14].

The token budget explains why the hierarchy exists. The tokenizer uses eight residual-vector-quantization codebooks, one with 16,384 entries for semantic content and seven with 1,024 entries for acoustic detail ^[10], and inference runs at 25 audio frames per second ^[11]. A five-minute song is therefore about 7,500 frames, or roughly 7,500 semantic tokens for the Global LLM and about 60,000 tokens in total across all eight codebooks ^[2]. Putting all of that on one autoregressive model would be an expensive way to lose the chorus.

The control surface is a real differentiator if it works. Users supply lyrics, which can carry section labels including [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo] and [Outro], plus a description that can specify genre, BPM, key, emotional progression, vocal timbre, backing vocals, instruments, percussion and arrangement changes ^[12]. That is a production vocabulary, not a single prompt box.

On the operational side, the model card says CPU offloading can bring video-memory use down to roughly 8 GB, while the repository instructions require CUDA, describe non-streaming generation and call for two GPUs ^[13]. Non-streaming matters: you wait for the whole five minutes before you hear anything.

The word to discount is "production-ready." MiniMax applies it to Music 3, and the support for that claim is currently company-selected samples and a reproducible generation example, not independent quality testing, disclosed customers or professional-studio adoption ^[5]. MiniMax's headline reach numbers, more than 300 million individuals across over 200 countries and more than 1 million enterprises and developers across over 100 countries, are company-wide and are not broken out for Music 3 ^[16].

Context on the balance sheet, because it shapes how long this kind of release stays free. MiniMax was founded in early 2022 by Yan Junjie and is based in Shanghai ^[15]. Alibaba led a $600 million round in March 2024 at a valuation above $2.5 billion, with HongShan participating, according to Fortune ^[17], and the company listed in Hong Kong in January 2026, raising about $619 million, according to Reuters ^[18]. Its 2025 annual report shows $79 million in revenue against $252.8 million in research and development, with 418 employees at year end ^[19]: R&D running about 3.2 times revenue ^[3] on revenue that grew about 2.6 times year over year ^[4].

What to watch: whether the checkpoint actually lands on the Hugging Face and ModelScope pages the announcement points to ^[3], what license arrives with it, and whether the two-GPU non-streaming path gets a streaming or single-GPU option. Until the weights exist, the useful output of this release is the architecture note.