An 8-billion-parameter language model and a 2.4-billion-parameter audio renderer sit behind MiniMax’s new music release, which turns a creative concept and optional lyrics into a complete, produced song in a single pass: Music 3.0, out August 13 with open weights, composes, arranges, performs, and produces a full track up to five minutes long, with vocals that sound performed rather than synthesized.
What MiniMax Announced
On August 13, MiniMax introduced Music 3.0, its next-generation music generation model, positioned as the strongest piece of its speech and music line alongside Speech 2.8. The pitch is different from most text-to-music tools. Instead of generating a generic loop or a short instrumental from a style label, Music 3.0 accepts a creative concept, like a scene, mood, or direction, plus optional lyrics, and then handles the entire production chain: composing, arranging, performing, and producing a complete song up to five minutes long in one pass.
MiniMax frames the release around the aspects of music creation that prompts usually cannot capture. The first is expressive intent, understanding what the creator actually wants to say. The second is sustaining that intent across a full-length song, which is where short-generation models fall apart. The third is rendering instruments with clarity and physical realism, and the fourth is generating vocals that sound like a performance, with breath, phrasing, and delivery, rather than a synthesized voice singing flat. Those four problems map directly to the architecture MiniMax chose to build.
The company published three demo directions with the release: a classic Shanghai jazz and soul piece with Mandarin lyrics, a melodic EDM track about digital identity, and a progressive house song built around regret and connection. The demos are deliberately varied in genre, language, and vocal character, and they show the model handling both Chinese and English vocals with different emotional registers.
How Music 3.0 Works
MiniMax redesigned the whole pipeline, from music description and language modeling to audio rendering, and the technical breakdown is unusually concrete. A fine-grained temporal description layer tracks how emotion, instrumentation, and vocal delivery evolve over time, instead of treating a song as a single static style tag.
The core is a hybrid architecture. An eight-layer residual vector quantization (RVQ) scheme represents music hierarchically: the first layer captures core semantics and structure, while the remaining seven layers progressively encode acoustic detail. The first layer is trained to learn a stable information backbone first, and then all codebooks are trained jointly, so the model does not overload a single token layer with both structural and fidelity information.
The language modeling side uses a hybrid of two models. An 8B “Global LLM,” initialized from Qwen3.5-8B, predicts semantic tokens frame by frame while modeling the whole-song context. A smaller 0.6B “Local LLM” predicts the within-frame acoustic tokens along the depth axis. Training happens in two stages: global alignment first, then full-parameter joint training of both models. The division of labor is structural stability from the big model, acoustic detail from the small one.
The most unusual decision is how audio gets rendered. Instead of feeding discrete acoustic tokens straight into a decoder, Music 3.0 fuses the continuous hidden states from both LLMs and uses them to condition a 2.4B flow-matching module, which feeds a 123M Flow-VAE that decodes the final audio. The full path is fused LLM features to flow matching to VAE hidden states to Flow-VAE decoder to audio. MiniMax says this preserves long-range consistency while improving pronunciation accuracy, instrumental coherence, and fine-detail fidelity. In total the generation stack is roughly 11B parameters, which puts self-hosting within reach of serious GPU owners but well beyond a laptop.
What the Architecture Buys
MiniMax summarizes the upgrades as three: more accurate interpretation of creative intent, more complete and varied arrangements, and clearer, more natural sound. The design choices map to each. The temporal description layer is what lets the model follow intent across a five-minute arc rather than a thirty-second loop. The hybrid global-local language modeling is what keeps the song structurally coherent, verses flowing into choruses with the right pacing, while the local model handles what actually happens inside each frame. And the flow-matching plus Flow-VAE path is what produces the physical realism, the sense that the drums were played and the piano was recorded in a room rather than synthesized in software.
For anyone who has used text-to-music tools, the practical difference is the length and the completeness. Most current models produce short clips, a verse, a hook, or a 30-second stem, and stitching them into a song is manual work. Music 3.0’s single-pass generation up to five minutes removes that stitching step, which changes the workflow for content creators who need a complete track with structure, dynamics, and an ending. The model also sits inside a wider MiniMax lineup, M3 for language, H3 for video, and Speech 2.8 for voice, that points toward a company building a full generation stack rather than point tools.
The field context matters too. Suno remains the best-known name in AI song generation, and the broader market includes UDIO and a growing list of specialized tools, which we round up in our guide to the best AI music generators of the year. MiniMax is entering as a later player with a heavier architectural bet: open weights, a full production pipeline, and a focus on realism that the consumer tools tend to trade away for convenience.
The Honest Caveats
Music quality is subjective, and the demos are curated. The published examples sound polished, but they are the company’s own selections, and real-world results with arbitrary prompts will vary more. The five-minute ceiling is generous for a song and restrictive for anything longer, like soundtrack work or concept albums, where creators still need composition or editing.
The open-weights framing comes with deployment reality. The full stack is roughly 11B parameters across the global LLM, local LLM, flow-matching module, and VAE, and MiniMax’s post does not detail the licensing or tooling story for the weights in the same depth it gives the architecture. Developers should check the actual license terms before building products on top of it, and self-hosting at useful speeds requires hardware most hobbyists do not have.
There is also the human-factor question that every singing model inherits. Vocals that “sound performed rather than synthesized” are the feature, but they also make the line between legitimate creation and voice cloning blurrier, and MiniMax does not address artist consent or voice rights in the announcement. Teams using the model commercially should think about that exposure themselves.
Who Should Care
Independent musicians and producers who want to iterate on complete song structures quickly are the clearest fit, especially if they write lyrics and want arrangement options rather than a single fixed interpretation. Content creators, YouTubers, and podcast producers who need full background tracks with real structure will save the most time, since the single-pass generation removes the stitch-it-yourself step. Developers and startups building music tools should watch the open weights closely: a production-grade, open music model changes what a small team can build. If you are a casual user who just wants a catchy loop for a video, the consumer tools remain simpler, and the extra realism here may not be worth the setup.
The Bottom Line
MiniMax Music 3.0 is the most technically ambitious music generation release of the year so far. A full-song, single-pass pipeline with open weights, hybrid language modeling, and flow-based audio rendering is a real step beyond clip generation, and the demos support the “performed, not synthesized” claim. The caveats, curated demos, deployment weight, unstated licensing details, and the voice-rights gray zone, are real but secondary. For anyone who needs complete songs rather than loops, this is the release to test first.