MiniMax-Music3: An Open-Weights Model for Full-Song Generation

What MiniMax-Music3 Actually Is

MiniMax-Music3 is an open-weights music generation model from MiniMax, released on Hugging Face. It generates complete songs up to five minutes long from two inputs: lyrics and a detailed music description. The output is 32 kHz, 16-bit stereo WAV audio.

The headline capability is long-range coherence. Rather than producing a short loop or a clip, the model is designed to hold a musical theme, rhythm, vocal identity, and arrangement together across a full song structure — intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro. Anyone who has tried to stitch together short AI music clips knows why that matters: keeping vocals and instrumentation consistent over minutes is the hard part.

Under the hood, MiniMax-Music3 uses a hierarchical setup. An 8B "Global LLM" (initialized from Qwen3-8B) handles long-range structure by predicting the first semantic codebook frame by frame. A smaller 0.6B "Local LLM" fills in the remaining acoustic detail per frame. Instead of decoding purely from discrete tokens, a synthesis path fuses the hidden states of both models and runs them through Flow Matching and a Flow-VAE decoder to produce the waveform. The stated aim is preserving richer detail for vocal articulation and instrumental texture.

The Control Layer Is the Interesting Part

For automation people, the most useful feature isn't the audio quality claim — it's the input format. MiniMax-Music3 takes two complementary, structured inputs:

Lyrics carry the words to be sung and can include explicit section tags like `[Intro]`, `[Verse]`, `[Chorus]`, `[Bridge]`, `[Instrumental]`, `[Solo]`, and `[Outro]`. Those tags map directly to song structure, so you can script the shape of a track programmatically.

The music description defines style, emotion, vocals, instrumentation, and production. The model card recommends a "Structured Caption" split into three sections: Global Metadata (genre, subgenre, BPM, key, scale, emotional progression, production profile), Vocal Details (gender, timbre, performance style, harmony, effects), and Arrangement (primary and secondary instruments, section-level evolution, groove, bass, percussion, textures).

That structure is what makes this practical to drive from code. A caption with named fields is something an LLM can generate reliably, which means you can build a pipeline: generate lyrics and a structured caption with a model like Claude or OpenAI, pass them to MiniMax-Music3, and get back a WAV. Predictable input schemas are far easier to automate than free-text prompts where you're guessing what the model will latch onto.

Running It and Where It Fits

MiniMax-Music3 is served through SGLang-Omni. The basic flow is to download the weights (`hf download MiniMaxAI/MiniMax-Music3`), then serve with `sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000`. Generation uses a shared speech-style API: lyrics go in the `input` field and the music description goes in `instructions`. The model card also lists diffusers and ComfyUI as inference options, and mentions a low-VRAM path and prompt enhancement, though the specifics of hardware requirements aren't fully spelled out in the material here.

One caveat worth noting: the diffusers code snippet on the page appears to be a generic image-generation example (it passes an image prompt and reads `.images[0]`), so treat that as boilerplate rather than a working music example and follow the SGLang instructions instead.

The obvious comparison is to hosted music services like Suno and Udio. The trade-off is familiar: those are polished, no-setup web products, while MiniMax-Music3 is open weights you host yourself. Self-hosting means no per-generation fees at scale, control over your own infrastructure, and the ability to wire the model into a private pipeline — but you own the GPU cost and the operational work. The structured caption format leans toward people who want repeatable, controllable output rather than one-off creative sessions.

For automation use cases, that's a real fit: background music for video generation, custom jingles or intros produced on demand, personalized tracks, or A/B variations generated in bulk. Because it exposes an HTTP endpoint, you can call it from a backend or slot it into a workflow tool like n8n, Make, or Zapier via an HTTP request node once the server is running.

As with any generative audio model, mind the legal and licensing side — check the model's license terms and how any generated output can be used commercially before building a product on top of it, since the page here doesn't settle those questions.

If you want help standing up MiniMax-Music3 or building a music-generation pipeline around it, browse the provider directory to find people who can put it to work.