What is MiniMax Music 3?
MiniMax Music 3 generates complete songs of up to five minutes with sung vocals from two inputs: lyrics with section tags such as [Verse] and [Chorus], and a written description of style, vocals and arrangement. It pairs an 8B global language model initialised from Qwen3-8B with a 0.6B local model for acoustic detail and a 2.4B flow-matching decoder, and outputs 32 kHz, 16-bit stereo WAV.
MiniMax published the weights on Hugging Face in August 2026 under its own MiniMax-Music3 Community License. It runs through SGLang-Omni, a diffusers pipeline or ComfyUI, and needs an NVIDIA GPU with CUDA.
The hosted version on the MiniMax API costs $0.15 per track of up to five minutes, but since 20 August 2026 the music and lyrics APIs have been closed to new users; only existing paying customers can call them.
Is MiniMax Music 3 worth it?
MiniMax Music 3 is the open-weight song model with the lightest commercial restrictions among the ones that sing well. Its licence allows commercial products with two conditions: show the name "MiniMax-Music3" in the product's interface, and ask MiniMax for written authorisation once yearly revenue passes $20 million. YuE2, which scores higher on quality benchmarks, is non-commercial.
It writes the song you give it. You supply the exact lyrics with section tags plus a description, and MiniMax recommends a three-part structured caption (genre, BPM and key; vocal details; arrangement); a published skill expands a short prompt into that format. On WildSongBench, a 192-prompt benchmark run by the YuE2 team, MiniMax Music 3 had a phoneme error rate of 6.27%, the lowest among the open models tested, so its sung words are the easiest to make out. Its SongBench average of 6.2830 trailed YuE2 at 6.7316 and Suno v5 at 6.8721.
The catch: hardware and access. The model card asks for a GPU with 24 GB or more of VRAM at full precision; CPU offloading brings that to about 22 GB, and streaming the language model layer by layer fits 8 GB at a slower speed. Inference requires CUDA, which rules out Apple Silicon and AMD cards. Output is 32 kHz, below the 44.1 kHz of Lyria 3.5 and the 48 kHz of YuE2. There is no streaming, and tempo, key and structure are steered rather than guaranteed. Anyone offering it as a service must build safeguards against infringing outputs, and the acceptable use policy requires disclosing machine-generated content posted publicly.
Our take: for a company that wants to host its own song generator and sell what it makes, this is the cleanest licence among the strong singers; a team without a 24 GB card, or one working on a laptop, will get further with ACE-Step 1.5.
Where it earns its place
A music-tech startup building a personalised birthday-song app
Generate a sung song with each customer's name and story in the lyrics
Fill a lyrics template with [Verse] and [Chorus] tags and render on a self-hosted 24 GB GPU through SGLang-Omni's /v1/audio/speech endpoint; the licence allows selling the songs as long as the app shows "MiniMax-Music3".
An advertising agency's in-house audio team
Turn approved jingle lyrics into several fully produced takes
Run MiniMax's caption-rewriter skill to expand the brief into BPM, key, vocal and arrangement details, then render takes with different seeds; the low phoneme error rate means fewer takes rejected for mumbled words.
A YouTuber running ComfyUI on an 8 GB GPU
Make full songs for videos locally, without a subscription
Load the ComfyUI workflow and enable layer-by-layer offloading; it is slower than a 24 GB card but costs nothing per song, and the licence's use policy requires labelling the music as machine-generated when posted.
Before you commit
- Setup
- Download the weights from Hugging Face, then serve with sgl-omni serve or load the diffusers modular pipeline, which installs diffusers from a pull-request commit until support is merged; a ComfyUI tutorial covers the node workflow.
- Integrations
- SGLang-Omni exposes an OpenAI-style /v1/audio/speech endpoint with lyrics in input and the description in instructions; diffusers ModularPipeline; ComfyUI.
- Data and privacy
- Self-hosted, so prompts and audio stay on your hardware. The model card does not describe the training data.
- Cost at scale
- No per-song fee when self-hosted; the cost is GPU time. The API's $0.15 per track applies only to accounts paying before 20 August 2026, and commercial use above $20 million annual revenue needs MiniMax's written authorisation.
- Limits
- Text prompt capped at 5,000 tokens and audio at 9,000 acoustic frames (25 frames a second); non-streaming only; CUDA required; requested tempo, key and structure may not be followed exactly.
What’s good
- Open weights, commercial use allowed up to $20 million yearly revenue
- Full songs up to five minutes with vocals from your own lyrics
- Lowest phoneme error rate (6.27%) among open models on WildSongBench
- Runs in SGLang-Omni, diffusers and ComfyUI; 8 GB cards work with layer streaming
- Structured captions steer BPM, key, vocals and arrangement
What’s not
- CUDA only; 24 GB of VRAM for full speed
- 32 kHz output, lower than rivals
- Hosted API closed to new users since 20 August 2026
- Commercial products must show "MiniMax-Music3" and run abuse safeguards
- No streaming; tempo and structure are not guaranteed
Best for
Companies that want to self-host a commercial song generator with clear vocals and can provide a 24 GB NVIDIA GPU.
MiniMax Music 3 alternatives
MiniMax Music 3: common questions
- Is MiniMax Music 3 worth it?
- MiniMax Music 3 sings your lyrics clearly for up to five minutes and lets you sell the result, provided you have a 24 GB NVIDIA card and put its name on your product. We rate it 4.2 out of 5.
- How much does MiniMax Music 3 cost?
- Weights free under the MiniMax-Music3 Community License, commercial use included; written authorisation needed above $20 million annual revenue. MiniMax API: $0.15 per track up to five minutes, existing paying users only since 20 August 2026.
- What are the downsides of MiniMax Music 3?
- CUDA only; 24 GB of VRAM for full speed; 32 kHz output, lower than rivals; hosted API closed to new users since 20 August 2026; commercial products must show "MiniMax-Music3" and run abuse safeguards; no streaming; tempo and structure are not guaranteed.
- Who is MiniMax Music 3 best for?
- Companies that want to self-host a commercial song generator with clear vocals and can provide a 24 GB NVIDIA GPU.
- How much does MiniMax Music 3 cost for a team or in production?
- No per-song fee when self-hosted; the cost is GPU time. The API's $0.15 per track applies only to accounts paying before 20 August 2026, and commercial use above $20 million annual revenue needs MiniMax's written authorisation.
- Is MiniMax Music 3 safe for company data?
- Self-hosted, so prompts and audio stay on your hardware. The model card does not describe the training data.
- What are the best MiniMax Music 3 alternatives?
- The closest alternatives we have reviewed are YuE2-3B, ACE-Step 1.5, Lyria 3.5 and Eleven Music v2.5.
Answers checked Nov 7, 2026.
Make MiniMax Music 3? Show this rating on your site
