Change one: Input side

Music 3.0 replaces a single sentence with a section-by-section spec sheet

Since coarse descriptions are the root cause, the fix is finer descriptions. Instead of using one global tag for the entire song, Music 3.0 takes a Structured Caption—a timeline-based, section-by-section specification broken into three parts:

How detailed is this spec sheet? The official demo page includes the full caption for each song—over 700 English words. This is a real excerpt from one of the urban soul tracks, and it's just a part of the "Vocal Detail" section:

Vocal Style: The delivery is marked by a conversational,
behind-the-beat phrasing... The bridge requires a dramatic
shift to a "whisper-singing" technique, utilizing heavy
vocal fry and breathy tail-offs to convey extreme proximity.

Harmony/Backing Vocals: The choruses are bolstered by tight,
three-part male harmony stacks... panned hard left and right
to envelop the lead.

Notice how its adjectives come with executable specifics: it doesn't just say "whisper-singing," it says to use it in the bridge; it doesn't just say "three-part male harmony," it says to pan them hard left and right. These aren't vague words like "moving" or "polished" that offer no direction. The model can follow along because this description is already a work order.

Change two: Structure

Square brackets in your lyrics are real structural commands

The spec sheet controls the substance of each section, but what defines the song's overall skeleton? The answer lies on the lyrics side.

When you write structural markers like [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], and [Outro] in your lyrics, the model treats them as structural instructions. They determine how many sections the song has, their order, and where instrumental breaks occur—without ever singing the word "Verse." This is the most common user error: Square brackets are executable; the lyric text conveys emotion, not structure. If you want the song to move from a verse to a bridge and back to a chorus, you lay that out in the lyrics yourself.

Companion tools

Officially open-sourced: 1,000 templates and an expansion skill

Most people can't write a 700-word professional spec sheet, which is the barrier with this input format. The official solution is an expansion system: you write a plain-language sentence, and it expands it into a full spec sheet using professional music terminology.

This "template library" isn't an abstract concept—it's right there in the GitHub repository: 1,000 plain-text template files, organized into 18 genre families, covering folk, EDM, hip-hop, jazz, metal, soul gospel, East Asian lyrical, country, and more. The expansion logic is also open-sourced as an Agent skill, installable with a single command:

npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter

Its retrieval method will look familiar to anyone who's worked with Agents: the skill explicitly forbids scanning all 1,000 templates, building a database, computing vectors, or calling external APIs. Instead, it narrows down progressively: first, it reads a genre routing table to identify one or two families; then it compares brief cards in the family index; finally, it fully reads only up to 3 selected templates, each serving a distinct purpose:

For those who don't plan to install the model but just want to write better prompts, these 1,000 templates are a readily usable library of arrangement prompt examples.

Mechanism: Core

Two models, two jobs: one remembers the whole song, the other focuses on each frame

We've covered the input side. But what inside the model actually supports a five-minute generation? First, let's understand how audio is stored in the model.

Audio can't be fed directly into a language model. It's first divided into discrete Audio Frames (here, 25 frames per second). Each frame is then encoded using an 8-layer codebook (officially called Residual Vector Quantization, or RVQ): Layer 1 uses a large codebook of 16,384 codewords and captures only the frame's core semantics and structure—the chords, the section, the words being sung. Layers 2 through 8 each use a smaller codebook of 1,024 codewords and progressively add acoustic detail residuals—timbre, overtones, spatial feel.

Two models: one that remembers the whole song, one that focuses on the current frame

Now that we know how sound is layered, the division of labor follows naturally:

Why split it this way? Asking an 8B-parameter model to manage both the song's overall direction and the overtone details of every frame simultaneously wastes resources and invites confusion—its attention gets stretched across two scales. By offloading detail work to a 0.6B model that handles frames one at a time, the larger model can focus on remembering what happened five minutes ago. This way, long-range stability and local richness are both preserved.

Official MiniMax Music 3.0 architecture diagram. From bottom to top, four layers: the bottom is the input (left purple box is Structured Caption, right green box is lyrics); above that is the Hybrid LM, where the dark blue horizontal strip is the Global Model. Its per-frame hidden states go up, pass through C₀ into the Local Model, which outputs C₁ to C₇ layer by layer. Two fusion arrows then feed the hidden states into the Synthesis layer, which uses Flow-Matching and a Flow VAE Decoder to produce audio. The red circle on the right is a stop token; the gray squares are hidden states. Source: MiniMax official launch blog.

Mechanism: Audio quality

The final step skips compression, preserving breaths and decay

But the step that truly determines whether a song sounds good is the last one—and it's also the easiest to overlook.

The conventional approach works like this: the language model emits a sequence of discrete codewords, and a decoder reconstructs the sound from them. The problem lies in the word "codeword"—they are quantized. A continuous waveform gets squeezed into a finite set of bins, and the subtle nuances in between (breaths, glissandos, the ring of a string after it's plucked) are lost at the moment of quantization. No amount of decoding afterward can recover them.

Music 3.0's approach is different: during inference, it doesn't load the discrete decoder at all. Instead, it takes the continuous hidden states from the final layers of both language models—the high-dimensional information before it's compressed into codewords—fuses them, and feeds them directly into a 2.4B-parameter flow matching module, which is then decoded into a waveform by a 123M-parameter Flow-VAE.

The Flow-VAE wasn't built from scratch; its architecture is adapted from MiniMax's own speech model, MiniMax Speech, and retrained for music's dynamic range and spectral characteristics.

Results

16 Official Demo Tracks, from Shanghai Jazz to Gospel Metal

Now that we've covered the mechanics, the remaining question is how it sounds—you'll have to listen for yourself. This is also the only evidence in this release: no benchmarks, no comparison experiments, just 16 demo tracks. The listening deck below lets you click through genre tags to hear different tracks. Each comes with its actual input description, showing how "the same model can produce wildly different results with a different input."

The last track's input specifically requests "a mature parental voice aged 50-70, with a Taiwanese accent, and no youthful bright tone." Whether such detailed requests about age and accent are fulfilled is the real test of whether this "Structured Caption" format actually works.

The Remaining Seven Tracks

The official blog releases a total of 16 audio tracks. Track 2 and Track 11 share the exact same input description and lyrics, leaving 15 unique songs; we've deduplicated them here. All audio files are direct links from MiniMax's official CDN.

License

Commercial Use Free up to $20M Annual Revenue

Sounds good, but can I actually use it? The release touts "open-source" in the title, yet the blog post itself never mentions the license once. So this entire section is sourced directly from the license text.

It's called the MiniMax-Music3 Community License. The full text is 7,400 bytes, and the opening wording closely mirrors the MIT License: free to use, modify, redistribute, and use commercially. There are only four additional conditions:

For the vast majority of individual creators and small teams, the only one requiring real action is the first: label the model name in your interface.

No Geographic Exclusion Clause

The same company released a very different license ten days earlier.

The end of the license also specifies the model's lineage: the Global Model is fine-tuned from Qwen3-8B (Apache 2.0), the diffusion part is adapted from Stability AI's stable-audio-tools (MIT), and the VAE is adapted from Descript's audio codec DAC (MIT).

Getting Started

Runs on 8GB VRAM, But the Official Model Card Lists Five Limitations

Since the license allows it, what about the hardware? Again, this part isn't in the blog post—it's all in the model card.

VRAM Tiers (For Single-GPU Setup)

First, let's be clear about what these tiers correspond to: running with diffusers on a single GPU. The officially recommended serving path has different requirements, detailed in the next section.

On the ComfyUI side, INT8 quantized model files are also provided specifically for smaller VRAM cards. For long track generation, you can also enable chunked decoding to further reduce VRAM usage.

Additionally, the HuggingFace model page already hosts three community fine-tunes and four quantized versions—if you'd rather not convert formats yourself, you can just download a ready-made one.

Three Deployment Routes

The Five Limitations Listed by the Official Model Card

This list is the vendor's own admission, which is more valuable than any marketing fluff:

This last point brings the whole discussion back to earth: the structured caption greatly increases the probability of "following instructions," but it's still probability, not the deterministic execution you'd get from feeding a score into a sampler.

Two Models: One Tracks the Song, One Handles Each Frame

Audio is sliced into 25 frames per second, each encoded into an 8-layer codebook: Layer 1 handles structure; the other 7 layers handle timbre details. These frames are processed sequentially by two differently-sized models.

Skip the Compression Step, Keep the Breaths and Slides

The step that makes or breaks the sound is the last one, and it's usually where quality gets lost.

Runs on 8GB VRAM, But the Model Card Lists Limits

The weights are open, the VRAM bar isn't too high, and commercial terms are permissive. The boundaries of this release are all spelled out in MiniMax's own model card.