MiniMax-Music3 Open Weights Are Out: What the Repo Shows and What Isn't Confirmed

MiniMax-Music3, the open-weights music generation model, appeared on Hugging Face on 2026-08-13 under the MiniMaxAI organization with no paper, no launch blog post, and no benchmark report — just the weights, a README, and a terse social post calling it the "next-generation, production-ready" text-to-music model. At the time of writing the repo had roughly 25 downloads and 60 likes, which is what a quiet release looks like. This is a what-we-know-so-far piece: everything below that is labeled "from the repo" comes straight from the model card and files, and everything that is not confirmed yet is marked as such.

First, a date correction

Two releases share this model family, and it helps to keep them apart. The hosted API model music-3.0 went live on MiniMax's platform on July 16, 2026, priced at $0.15 per generated piece of up to five minutes, with a rate-limited free tier. The open-weights release under the name MiniMax-Music3 landed on Hugging Face almost a month later, on 2026-08-13. The vendor has not said explicitly whether the open checkpoint is the same weights that power the API — that equivalence is one of the unconfirmed items on this list — but the README describes the same five-minute full-song capability and the same structured-caption control surface, so they are at minimum the same generation.

What the repository actually contains

MiniMax-Music3 is a hybrid hierarchical architecture, and the README is unusually specific about the pieces. A Global LLM of 8B parameters — initialized from Qwen3-8B — predicts the first residual vector quantization codebook frame by frame and carries the song's long-range structure: themes, rhythm, vocal identity, arrangement progression. A Local LLM of 0.6B parameters predicts the remaining acoustic codebooks within each frame. Instead of decoding purely from discrete tokens, a Flow Matching module (2.4B parameters) fuses the final hidden states of both LLMs and synthesizes a continuous latent, which a Flow-VAE decoder (123M parameters) turns into 32 kHz, 16-bit stereo audio.

The tokenizer is an eight-layer RVQ stack: the first semantic codebook holds 16,384 entries; the seven acoustic codebooks hold 1,024 each. Training first optimizes the semantic codebook, then all eight jointly.

What ships in the repo, concretely: the language model weights in four safetensors shards, a 48-shard Qwen-based backbone under qwen_7B/, the Flow-VAE (flowmatching_vae.pth), the RVQ depth decoder, condition encoder, vocoder, scheduler, tokenizer files, a diffusers modular-model index, and an end-to-end test script with reference audio. That is a runnable model, not a weights-in-a-box teaser.

The control surface is the interesting part

The input format is what separates this from prompt-and-pray music generation. MiniMax-Music3 takes lyrics and a music description, and the README recommends a structured caption with three sections: global metadata (genre, subgenre, BPM, key, scale, emotional progression, listening scenario, production profile), vocal details (gender, timbre, performance style, harmony, backing vocals, effects), and arrangement (primary and secondary instruments, section-level instrument evolution, groove, bass, percussion, texture, spatial effects).

Lyrics accept explicit section tags — [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo], [Outro] — so the model can be told exactly where the chorus lands and what happens in the bridge. That is a level of structural control that most consumer music generators expose only through post-hoc editing, if at all. The hosted API adds the same controls as parameters: an is_instrumental flag, a lyrics_optimizer for auto-lyrics, and an audio setting covering MP3, WAV, or PCM at 16/24/32/44.1 kHz and 32/64/128/256 kbps.

What it costs to run yourself

This is not a heavy model by 2026 standards. The README states full precision fits in 24 GB of VRAM, generation runs in roughly 22 GB with automatic CPU offloading, and streaming the language model layer by layer brings it down to fit 8 GB cards. Three serving paths are documented: SGLang-Omni (a one-line sgl-omni serve command), a diffusers modular pipeline (the integration PR is huggingface/diffusers#14456, still unmerged at the time of writing), and ComfyUI.

That serving story matters more than the architecture for most readers, because it is the difference between "a model that exists" and "a model you can actually operate." The diffusers path is a single from_pretrained call and produces a 32 kHz stereo WAV file you own, on your own disk, with no rate limit and no per-song meter.

The license that actually attaches

MiniMax-Music3 ships under a "MiniMax-Music3 Community License," which is permissive in the practical sense but has strings. You may use, copy, modify, and sublicense the software and weights; you must display "MiniMax-Music3" prominently in the UI of any commercial product that uses it; and if aggregate yearly revenue from such products exceeds $20 million USD, you need separate written authorization from MiniMax. There is also an Acceptable Use Policy (Exhibit A) that covers the usual prohibitions plus disclosure requirements for machine-generated content.

Underneath it, the model is a stack of borrowed parts with more permissive origins: the Global LLM is fine-tuned from Qwen3-8B (Apache 2.0), the DiT-2B flow component is modified from Stable Audio's code (MIT), and the VAE is modified from the DAC audio codec (MIT). So the community-license restrictions sit on top of Apache/MIT components, and the attribution requirement is the practical constraint most commercial users will feel.

The hosted route that shipped first

If you do not want to operate weights at all, the API version has been live since July 16. Model IDs music-3.0 (paid, 120 requests per minute) and music-3.0-free (free, 3 RPM) sit behind api.minimax.io/v1/music_generation, with a separate lyrics-generation endpoint at $0.01 per song. Output URLs expire after 24 hours, or audio can be returned as hex for streaming — a design clearly built for app integration rather than archival.

Pricing for the API is flat per piece: $0.15 per generated song up to five minutes. The free tier is the same model at 3 RPM, which is genuinely usable for prototyping. If you need self-hosted capacity instead, the open weights change the cost equation entirely — you pay for GPUs, not per song, and the per-song meter disappears.

What is not confirmed

Being precise about what we do not know is the point of this piece.

• No independent evaluation exists. Every capability claim above comes from MiniMax's own README and API docs. No third party has published a listening test, a music-quality benchmark, or a head-to-head against Suno, Udio, or Google's Lyria 3.5 as of writing.

• Open checkpoint vs API checkpoint. The vendor has not stated whether the weights on Hugging Face are the exact model behind music-3.0. Assume "same generation" until told otherwise, but it is unconfirmed.

• No paper and no training-data statement. There is no technical report, and the README does not describe the training corpus, its licensing, or a provenance audit — a live question for anyone shipping commercial music, where the training-data posture of closed rivals has been the subject of multi-year litigation.

• No self-hosting cost guidance. The 22 GB figure is memory, not price. Nobody has published tokens-per-song or an hourly generation throughput for this model on typical hardware.

Why this matters to builders

A two-day-old checkpoint has not been load-tested by anyone but the lab, and that is exactly the situation where routing with automatic failover earns its keep. A routing layer — the kind OrcaRouter runs for 200+ models at provider-list-price pass-through — lets you point a test path at a fresh open-weights model and fall back to a proven one when it stalls, without a second contract or a code change. You can treat MiniMax-Music3 as an experiment this week and a dependency next month, and the plumbing does not have to change between those two states.

The strategic read is simpler. MiniMax is now one of the few music-generation vendors that offer both a paid API and open weights for the same product family. Suno and Udio are closed platforms. Google's Lyria 3.5 is a consumer-app feature. If the open checkpoint is even close to the API version in quality, the five-minute length, the structured-caption control, and the 22 GB serving footprint make it a credible self-hosted foundation — which is a category that did not have a credible entrant a week ago.

What to watch next

The two things that would move this story from "interesting repo" to "durable platform": the first independent listening tests (they tend to appear within days of an open music-model release), and whether the diffusers integration merges, which would put MiniMax-Music3 into every existing diffusers music pipeline. Watch also for a paper, a training-data statement, and any change to the hosted API price — on a pass-through router, a vendor price cut is live the same day, which is the moment the hosted route becomes the obvious default for most teams.