MiniMax Music 3

Two weeks ago we wrote up YuE2, the open music model that scores highest and then tells companies they cannot use the weights. MiniMax Music 3 is the other side of that trade. It scores lower, it sounds a step behind the current Suno releases, and it is licensed so you can put it in a product and sell it. If you are building something rather than benchmarking something, that is the one that matters.

First, the name, because MiniMax made this hard

MiniMax Music 3 is not MiniMax M3. We covered MiniMax M3 earlier this year: that is the 1M-token sparse-attention coding model. This is a text-to-music model with a completely different architecture, a different license, and a different repository. The weights even live under two different Hugging Face org spellings, MiniMaxAI on Hugging Face and MiniMax-AI on GitHub. Check which one you are downloading before you pull 57 GB.

The license is the story

MiniMax Music 3 ships under the MiniMax-Music3 Community License, which is shaped like the Qwen and Llama community licenses rather than a research license. Commercial use is granted up front. Fine-tunes, merges and derivative works are permitted, in the license's words the right to "use, copy, modify, merge, publish, distribute, sublicense, and/or provide copies".

Two conditions come with it. If you ship a commercial product built on this, you "shall prominently display 'MiniMax-Music3' on the user interface". And if your yearly revenue from that product passes 20 million US dollars, you need "a separate, prior written authorization from MiniMax", requested by email. There is also an output-disclosure duty: material you publish has to be clearly marked as machine-generated.

Compare that to what else is on the table. YuE2's weights are CC BY-NC 4.0 with a carve-out that lets individuals sell songs but explicitly excludes company use. ACE-Step 1.5 is MIT and has no conditions at all. MiniMax sits between them: a real commercial grant with an attribution string and a revenue ceiling that almost nobody reading this will hit. For a solo developer or a small studio, the practical answer is that you can ship it, you just have to say so in your UI.

What is inside

The pipeline has three stages and they are worth understanding, because they explain both the quality and the hardware bill.

The tokenizer is an eight-layer residual vector quantizer. The first layer carries semantics and structure with a 16,384-entry codebook, and layers two through eight each add acoustic detail with a 1,024-entry codebook. On top of that sits what MiniMax calls a Hybrid-LM: an 8B global language model initialized from Qwen3.5-8B handles long-range song structure, and a separate 0.6B local model, randomly initialized, handles frame-level acoustic detail.

Those fused features then condition a 2.4B flow-matching module, and a 123M Flow-VAE decoder turns the result into audio. Output is 32 kHz, 16-bit stereo WAV, up to about five minutes in a single generation. That is roughly 11B parameters of model across the stack, and the published checkpoint weighs 53.4 GiB on disk, or about 57 GB.

The design choice worth noting is that this model writes a song forwards, frame by frame, rather than denoising a whole track at once the way ACE-Step does. That is why it holds five-minute structure well and why it is slower than the diffusion-first models.

Where it actually lands

MiniMax published no benchmark numbers with the release, which is its own kind of answer. The most useful figures come from a competitor: M-A-P's WildSongBench, run as part of the YuE2 release, included MiniMax Music 3 and measured it locally.

Read the second column, not the first. MiniMax Music 3 has the lowest phoneme error rate of any open model in that table at 6.27%, which means it pronounces lyrics more reliably than the model that beat it overall. If your use case is songs with words that people need to understand, that number is the one that predicts whether you will be happy. These are author-reported figures from M-A-P, and a competitor scoring you well is a better signal than scoring yourself.

Listener reactions are more sober than the scores. The recurring community read is that Music 3 lands somewhere between Suno v3.5 and v4.5, which is to say clearly behind the current commercial state of the art and clearly ahead of where open music generation was a year ago. One of the top threads on the model card is titled, bluntly, that the quality looks like Suno V3.5. Set expectations there rather than at the benchmark table.

What it takes to run

The reference path is the expensive one. MiniMax's own server setup splits the model across two CUDA GPUs: one runs the language model and the eight-codebook autoregressive generation, the other runs flow matching and waveform decoding. That is the configuration the README documents, and it is not a consumer setup.

So the headline requirement is two GPUs, and the real requirement is one 24 GB card, or an 8 GB card if you accept group offloading and the speed penalty that comes with it. The community did the usual work fast: there are GGUF builds, MLX ports for Apple Silicon and 4-bit quantizations, and the ComfyUI repack has been downloaded roughly 37 times more often than MiniMax's own checkpoint. That ratio tells you how people actually run this.

Run it today

Pull the weights, start the server, post a request. The API is OpenAI-shaped, which makes it easy to drop into something you already have.

hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm

sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000

curl http://127.0.0.1:8000/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "minimax_ttm",
    "input": "[Verse] lyrics go here",
    "instructions": "indie pop, 96 BPM, female vocal, warm analog production",
    "response_format": "wav",
    "seed": 7,
    "max_new_tokens": 9000,
    "stream": false
  }'

Two fields do the work. The input carries lyrics with section tags: Intro, Verse, Pre-Chorus, Chorus, Post-Chorus, Bridge, Instrumental, Solo and Outro. The instructions field carries a structured caption covering global metadata (genre, BPM, key, emotional arc), vocal details (gender, timbre, performance style) and arrangement (instruments, how sections evolve, effects). Writing that caption well is most of the skill, and MiniMax evidently knows it, because the repository ships a music-caption-rewriter agent skill whose whole job is turning a vague description into a structured one.

That is the second open music model in a row to ship a first-party agent skill, after YuE2. It is becoming a convention in this corner of the ecosystem, and it is a reasonable one: these models have a text front door that rewards careful prompting.

If you would rather not run a server, ComfyUI supports it natively with FP16 and INT8 builds plus a separate text encoder and VAE, and that is the fastest way to hear it on your own machine.

The limits, stated plainly

Inference requires CUDA, so there is no supported AMD or CPU path from MiniMax itself, though the community GGUF and MLX builds exist. Only non-streaming generation works today, meaning you wait for the whole song. The text prompt caps at 5,000 tokens and audio generation caps at 9,000 acoustic frames, which is what produces the five-minute ceiling.

The documentation is also honest that section tags and captions "provide generative control rather than strict symbolic guarantees", so requested tempo, key, instrumentation and structure are steering, not contracts. There is no published fine-tuning procedure, so adapting it to a specific voice or genre means prompt engineering or community tooling. Language support is undocumented: the examples are English throughout, and while the Qwen3.5-8B backbone is multilingual, nobody has published results for tonal or non-Latin lyrics.

MiniMax Music 3 or YuE2

Pick MiniMax Music 3 if the output is going into something you sell, if lyric intelligibility is the thing you cannot compromise on, or if you want an OpenAI-shaped endpoint you can point existing code at. The attribution string in your UI is a cheap price for a clean commercial grant.

Pick YuE2 if you want the highest quality available in the open, if you want the editable symbolic score it generates before the audio, or if you are working alone and never plan to incorporate. Pick ACE-Step 1.5 if you want MIT with no conditions, fast iteration and a sub-4 GB footprint. And if you want a whole studio around the model rather than the model alone, our local AI music studio coverage tracks theDAW, with part two on the migration after it was renamed.

Sources and further reading

• MiniMax-Music3 on GitHub
• Model card and weights on Hugging Face
• MiniMax-Music3 Community License, full text
• MiniMax's own architecture write-up
• ComfyUI tutorial and model file layout
• Comfy-Org repack, FP16 and INT8 builds
• GGUF quantizations
• The music-caption-rewriter agent skill
• WildSongBench results, where the comparison numbers come from

Ten minutes, one decision: open the license, read the two conditions, and work out whether an attribution string in your interface is acceptable for what you are building. That question decides between this model and YuE2 faster than any listening test will.

Tested on: not independently tested. Architecture figures, audio specification, hardware paths and limitations come from MiniMax's published README, model card and architecture write-up. The benchmark figures are author-reported by M-A-P on WildSongBench, not by MiniMax and not by us. Quality impressions are community-reported. Download counts and repository statistics were read from the GitHub and Hugging Face APIs on the date below.
Date checked: 2026-09-20