MiniMax Music 3: open-weights model writes full songs up to five minutes long

What's going on

MiniMax is a Shanghai-based AI company and one of China's heavily funded "AI tigers." With MiniMax Music 3 (MiniMax-Music3 on Hugging Face), it has released a music model that writes and sings complete songs. You provide two things: lyrics with the sections marked – verse, chorus, bridge and so on – and a description of how the music should sound. What comes out is a song of up to five minutes in 32 kHz stereo, with vocals, instruments and transitions between sections.

The catch – in a good way: you can download the weights for free. According to OrcaRouter's blog, they appeared on Hugging Face on August 13, 2026, with no paper and no launch blog post – just a README and a short announcement in which MiniMax calls the model "production-ready." A month earlier, on July 16, 2026, the paid API version music-3.0 had already launched. AlphaSignal has now covered the model in depth and points to ready-made packages for ComfyUI, a popular graphical interface where you build AI workflows by snapping together nodes.

How the model works: plan first, then polish

Music 3 uses a hierarchical autoregressive architecture. Autoregressive means the model generates the song piece by piece, with each new piece building on the ones before. Hierarchical means several sub-models work on different timescales. Think of it like a studio: a producer plans the whole song, while engineers take care of how every single second sounds.

• Global LLM (8B parameters): Based on the Qwen3-8B language model, it keeps the whole song in view – lyrics, themes, rhythm, voice and arrangement. Technically, it predicts the first of eight codebooks frame by frame.
• Local LLM (0.6B parameters): Within each time step, it fills in the remaining seven codebooks with fine sonic detail such as timbre and texture.
• Flow matching (2.4B parameters): This module combines the internal states of both language models and generates a continuous, compressed audio representation section by section. Flow matching is a technique that gradually turns noise into a meaningful signal, similar to how image generators work.
• Flow VAE decoder (123M parameters): It turns that compressed representation into audible 32 kHz, 16-bit stereo audio. According to Intelligent Living, it comes out of MiniMax's work on speech synthesis.

During training, an RVQ tokenizer breaks the audio into building blocks. RVQ stands for residual vector quantization: the sound is encoded in several layers, from coarse to fine – a bit like sketching the outlines first and then adding detail. The first, "semantic" codebook has 16,384 entries; the seven acoustic ones have 1,024 each. During generation, according to Intelligent Living, the model bypasses the discrete decoder entirely and works directly with the combined states of the language models. All told, the system comes to around 11 billion parameters.

OrcaRouter notes that several components come from more permissively licensed projects: the global LLM from Qwen3-8B (Apache 2.0), the flow component from Stable Audio's code (MIT) and the VAE from the DAC audio codec (MIT).

Lots of control over your song

What mainly sets Music 3 apart from many music generators is the input. In the lyrics, you can mark sections like [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo] and [Outro], deciding exactly where the chorus lands and what happens in the bridge. For the music description, the README recommends a three-part structure:

1. Global settings: genre, subgenre, tempo (BPM), key, scale, mood arc, listening context, production style
2. Vocals: gender, timbre, delivery style, harmonies, backing vocals, effects
3. Arrangement: lead and supporting instruments, how the instrumentation changes per section, groove, bass, drums, texture, spatial effects

According to Intelligent Living, a bundled prompt-rewriting tool expands short keywords into the detailed format the model prefers. Users particularly praise how verse, chorus and bridge stay recognizable across five minutes and how stable the voice sounds.

How good is it? Only impressions so far

There are hardly any independent benchmarks. AlphaSignal puts the quality roughly at Suno V3.5 level: strong on vocals and lyric timing, but weaker than Suno V5 on instrumental texture. According to OrcaRouter, at the time of its report nobody had published listening tests or head-to-head comparisons with Suno, Udio or Google's Lyria 3.5. So every capability claim comes from MiniMax itself.

Intelligent Living compares Music 3 with its open rival ACE-Step 1.5:

By that measure, ACE-Step is leaner, faster, more permissively licensed and offers editing features like covers, stem separation and song extension. Music 3 can only generate new songs. Its advantage, according to the review, is the large 8B planner that holds long songs together, plus precise control. Intelligent Living reports barely any audible artifacts.

What you need to try it

By today's standards, the model isn't overly heavy. According to the README:

• 24 GB VRAM for full precision, e.g. an RTX 3090 or 4090
• approx. 22 GB VRAM if parts are automatically moved to system RAM (CPU offloading)
• 8 GB VRAM if the language model is loaded layer by layer (group offloading). That's slower and needs plenty of system RAM.

Officially, only Nvidia GPUs (CUDA) are supported. For AMD (ROCm) and Apple machines there are community ports, including MLX versions in MXFP4 format for Apple silicon and GGUF quantizations. Quantization means storing the model's numbers at lower precision so it needs less memory. There are three official ways to run the model: ComfyUI (via a ready-made package from Comfy-Org), a Python pipeline with Diffusers, and the SGLang-Omni server. According to OrcaRouter, the Diffusers integration hadn't been officially merged at the time of its report (pull request #14456). MiniMax hasn't said how long a song takes to generate. Budget several dozen gigabytes of disk space for the files.

License and hosted alternative

The MiniMax-Music3 Community License lets you use, copy, modify and sublicense the software and weights. If you use the model in a commercial product, you have to display "MiniMax-Music3" prominently in the user interface. Once such products bring in more than $20 million a year, you need separate written permission. An acceptable use policy bans, among other things, military applications, harmful content and impersonating people, and requires disclosing machine-generated content.

If you'd rather not run your own hardware, there's the API. music-3.0 costs $0.15 per song of up to five minutes and allows 120 requests per minute. The free tier, music-3.0-free, allows three requests per minute. A separate lyrics endpoint costs $0.01. The API returns MP3, WAV or PCM at up to 44.1 kHz. According to OrcaRouter, MiniMax hasn't confirmed whether the open weights are identical to the API model.

Interest is high. Intelligent Living counted more than 14,000 downloads and over 1,000 likes in the first week. According to AlphaSignal, the ComfyUI package's counter most recently stood at more than 950,000 – though that counts every file request, not every individual install.

Our take

Music 3 is the first music model from a major provider that's available both as a paid API and with open weights. Suno and Udio are closed platforms, and Google's Lyria lives inside consumer apps. If you generate songs yourself, you only pay for electricity, have no per-song limit and keep your data on your own machine.

There are caveats, though. Independent quality measurements are missing, and a Suno V3.5-level rating lags behind today's commercial competition. The model mostly handles English and offers neither editing features nor LoRA training. The license requires attribution for commercial use. Most sensitive of all: MiniMax says nothing about its training data. With music AI, that's no side issue – the training data of closed competitors has been fought over in court for years. If you plan to use AI songs commercially, keep that in mind.

MiniMax is betting on open weights in video, too – for example with its H3 video model, which has a similar $20 million revenue threshold. According to eWeek, though, key components there remain proprietary. For Music 3, by contrast, OrcaRouter says the repository contains a fully working pipeline.