MiniMax Music 3.0 Puts a Song Model on Your Own Hardware. Read the License First.

MiniMax Music 3 is an open-weight music generation model from the Chinese lab MiniMax, published on August 13, 2026. Give it a concept and optional lyrics and it composes, arranges, performs, and produces a complete song of up to five minutes. Unlike every commercial music generator, you can download it and run it on a single graphics card.

That is the headline, and it is real. The part worth reading carefully is the license, because "open weights" and "open source" are not the same thing, and the difference here has a dollar figure attached.

A song model you can actually download

Almost every music generator a business operator has heard of is a hosted service. You subscribe, you generate, and the model stays on someone else’s hardware. Our companion piece on Suno covers what that arrangement looks like from the inside, including the download caps arriving on that platform in September.

MiniMax Music 3 inverts the arrangement. MiniMax published the weights to Hugging Face, so the model runs where you put it. Nothing meters your output, nothing caps your downloads, and no terms-of-service change next quarter can reduce what you are allowed to take off the platform, because there is no platform. That is a structurally different relationship with a tool, and for anyone producing audio at volume it is the whole argument.

MiniMax is the same lab behind the M-series language models. Our explainer on MiniMax M3 covers the company and its release pattern, which has been consistent: capable models, published weights, and a license that repays reading.

What is inside it

The announcement and model card describe a hybrid system rather than a single network. A global language model of about 8 billion parameters, initialized from Qwen3.5-8B, handles long-range structure: verses, choruses, how the song holds together over minutes. A smaller local model of about 0.6 billion parameters handles frame-level acoustic detail. A flow-matching module of about 2.4 billion parameters and a 123-million-parameter Flow-VAE decoder turn tokens back into audio, working from an eight-layer residual vector quantization tokenizer.

The split is the interesting design choice. Music has structure at two very different timescales, and most earlier attempts either lost the shape of a song over three minutes or produced coherent structure with muddy audio. Handing those jobs to different models is a reasonable answer.

Generation limits are concrete: complete songs up to five minutes, a ceiling of 9,000 acoustic frames, and text prompts up to 5,000 tokens. MiniMax published no benchmark scores and no listening-test results, which is worth stating plainly rather than filling the gap with third-party claims.

What it takes to run

Inference requires CUDA, which rules out Apple silicon for the reference implementation, though community ports exist. The model card gives three working points: full precision fits under 24 GB of VRAM, CPU offloading brings it to about 22 GB, and a streaming approach fits on 8 GB cards.

That last figure matters more than it looks. An 8 GB card is a five-year-old consumer GPU. A model that writes and performs a full song on hardware that old is a meaningfully different proposition from one needing a workstation, and it is the reason the release landed in ComfyUI workflows within days, alongside community GGUF and Apple silicon builds.

What none of that tells you is speed. MiniMax has published no generation times, and the streaming path that fits an 8 GB card is trading memory for latency rather than getting something for nothing. Budget for a real test rather than assuming a five-minute song arrives in anything like five minutes.

The license is not open source

Here is where care is required. MiniMax Music 3 ships under the MiniMax-Music3 Community License, not Apache 2.0, not MIT. The distinction is not pedantry, because two clauses have direct commercial consequences.

The license text grants broad rights to use, copy, modify, merge, publish, distribute, and sublicense the software, including the weights, free of charge. Then it attaches conditions:

For a small studio, an agency, or an independent creator, none of that bites. Display the name, stay under twenty million, done. For a company above that revenue line, the weights being downloadable does not mean the model is free to deploy, and the authorization request goes to a MiniMax email address with no published turnaround or terms.

Note also what the license does not say. It grants rights in the model, not in the training data, and it makes no representation about the provenance of the music the model learned from. Anyone who has followed the litigation around hosted music generators should read that silence as unresolved rather than settled.

Compare that with Meta’s Muse Glimmer, released three days earlier under Apache 2.0 with no revenue trigger and no attribution requirement. Both are described as open weights in coverage. Only one is open source. Our piece on open versus closed AI models covers why this distinction usually matters more than whether the weights are downloadable.

The carve-out that is not there this time

There is one genuinely good piece of news in the license, and it is easy to miss because it is an absence.

Eleven days earlier, MiniMax released its H3 video model under a community license that excluded the United States, the European Union, the United Kingdom, and South Korea from local deployment. Anyone in those markets could read about the model but not lawfully run it. That carve-out drew significant criticism.

The Music 3 license, effective August 6, 2026, contains no territorial definition and names no excluded countries. Western users can deploy it. Given how recent the H3 restriction was, this looks like a deliberate correction rather than an oversight, and it should be read as the more meaningful signal about where MiniMax is heading than any architectural detail.

MiniMax Music 3 against Suno

These two are the clearest open-versus-closed comparison in generative audio right now, and they fail in opposite directions.

Neither has published a credible head-to-head evaluation, so anyone claiming one is better is telling you about their ears, not about the models. The honest position is that Suno is a product and MiniMax Music 3 is a component, and the right answer depends on whether you want a tool or an ingredient.

Who this is for

Run MiniMax Music 3 if you generate audio at volume, if download limits on a hosted platform would break your workflow, if you need the audio produced inside your own infrastructure for client or data reasons, or if you want a model that cannot be retired out from under you. It is also the sensible choice for anyone building music generation into a product, given that the attribution requirement is a line of interface text rather than a revenue share.

Skip it if you want to make a song this afternoon without installing anything, if you are above the twenty-million revenue line and cannot wait on an authorization email, or if you have no CUDA hardware and no appetite for community ports.

The practical first step is cheap: pull the model, generate against your own reference prompts, and compare the output with whatever you use today before committing to either hardware or a subscription. Nobody has published numbers that will make that decision for you.

Frequently Asked Questions