MiniMax-Music3-GGUF Full-Length Music Generation Model: 12GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | ggml-org/MiniMax-Music3-GGUF |
| Publisher guide | MiniMax: models and licenses |
| Published | 2026-11-17 |
| License | minimax-music3-community-license |
| Formats | GGUF |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
ggml-org has released a GGUF quantized version repository, ggml-org/MiniMax-Music3-GGUF, for the music generation model “MiniMax-Music3" developed by MiniMaxAI. This model is a text-to-audio generation model that outputs songs complete with vocals and accompaniment based on text lyrics and musical style instruction prompts.
With this GGUF version, you can generate music from your terminal or a local server using the llama-tts command or the /tts endpoint of llama serve in the llama.cpp project.
Note that much of the information regarding the architecture and generation quality introduced in this article is based on the official model card of the source model, MiniMaxAI/MiniMax-Music3.
Specifications
The specifications of the source model and the GGUF converted version are as follows:
- Architecture (Source Model): Hierarchical Autoregressive Model (Hybrid-LM) and Continuous Hidden-State Synthesis
- Global LLM: 8B parameters (initialized from Qwen3-8B; models the overall structure and semantic progression of the song) – Local LLM: 0.6B parameters (models frame-level acoustic details) – Synthesis: 2.4B parameter Flow Matching and 123M parameter Flow-VAE Decoder
- Tokenizer: 8-layer Residual Vector Quantization (RVQ). The 1st layer (semantic) has 16,384 entries, and the remaining 7 layers (acoustic) have 1,024 entries each
- Output Specifications:
- Standard output of source model: 32 kHz, 16-bit stereo WAV audio – Generation specification note for GGUF version: 44.1 kHz stereo audio (vocals and accompaniment) – Maximum generation length: Up to 5 minutes (9,000 acoustic frames at 25 fps equivalent)
- Input Formats:
[Lyrics]structure (supports section tags such as[Intro],[Verse],[Pre-Chorus],[Chorus],[Bridge],[Outro], etc.) –[Tags]or instructions (structured captions for genre, BPM, key, vocal expression, instrumentation, etc.) – Text prompt limit: Up to 5,000 tokens- License Terms: Distributed under the MiniMax-Music3 Community License.
Performance and Quality
The model card for the source model MiniMaxAI/MiniMax-Music3 publishes the following comparison and verification table as reproduction examples for playback and generation.
| Use case | Request | Result |
|---|---|---|
| Text-to-music | View script | minimax_ttm.wav |
According to the model card, a major feature of this model is its ability to generate entire songs lasting up to 5 minutes while maintaining overall consistency. From the intro to the verse, chorus, bridge, and outro, it can maintain themes, rhythms, vocal identity, and arrangement progression without breaking down. Additionally, the description explains that the synthesis method, which fuses the hidden states of the Global LLM and Local LLM and passes them to Flow Matching, achieves high quality in clear vocal pronunciation and instrument texture expression.
On the otherانه limitations, control via section tags and musical descriptions functions as generative guidance, and the generated tempo, key, specified instrumentation, lyrics, and song structure may not always strictly match the requested details.
Strengths and Use Cases
MiniMax-Music3 excels at generating full-length songs (up to 5 minutes) combining vocals and accompaniment from input lyrics and musical style instruction prompts. Unlike conventional short-form models, it can maintain a consistent song structure throughout the entire track without breaking themes, rhythms, vocal timbre, or arrangement progression.
With this model, you can meticulously control the song’s progression by inserting specific structure tags into the lyric text. The main available structure tags and the roles of input parameters are as follows:
- Lyric Section Tags: By placing tags such as
[Intro],[Verse],[Pre-Chorus],[Chorus],[Post-Chorus],[Bridge],[Instrumental],[Solo], and[Outro]on each line, you can directly specify changes in song paragraphs. - Global Metadata: You can specify genres, subgenres, BPM, keys, scales, emotional progression, listening scenarios, production profiles, and more.
- Vocal Configuration: You can specify gender, voice quality, singing style, harmony or backing vocal composition, and vocal effects.
- Arrangement Configuration: You can fine-tune primary and secondary instruments, instrumental changes across sections, grooves, basslines, percussion, spatial effects, and more.
Additionally, the music-caption-rewriter skill is officially provided as a prompt assistance tool that automatically expands short natural language descriptions into such structured captions (Global Metadata, Vocal Details, Arrangement). Using this, you can create expressive instructions while preserving the main body of the lyrics.
The newly introduced GGUF version (ggml-org/MiniMax-Music3-GGUF) is well-suited for easily trying out these advanced music generation features locally in your terminal or via a simple HTTP server leveraging the llama.cpp ecosystem.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 2.4B parameters (taken from the base model MiniMaxAI/MiniMax-Music3)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 12GB (RTX 4070 / 3060 12GB, etc.) | Q8_0 | 8.5GB | 10.2GB |
| 24GB (RTX 4090 / 3090, etc.) | BF16 | 16.0GB | 19.2GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
The publisher distributes this model as GGUF.
License — minimax-music3-community-license: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (minimax-music3-community-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Distributed Files
Weight files published in ggml-org/MiniMax-Music3-GGUF, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
MiniMax-Music3-BF16.gguf |
17.18GB |
MiniMax-Music3-Q8_0.gguf |
9.13GB |
mmproj-MiniMax-Music3-BF16.gguf |
6.64GB |
How to Get It
The GGUF converted model is available on the Hugging Face repository ggml-org/MiniMax-Music3-GGUF. Prior license agreement procedures (gated) are not required for use.
This model is automatically converted using ggml-org/convert and can be used by directly calling it from command-line tools or server features included in llama.cpp.
To generate audio files directly from the terminal, run the following command:
llama-tts -hf ggml-org/MiniMax-Music3-GGUF -f song.txt --output out.wav
AAlternatively, to start a local server and generate via the /tts endpoint, launch it with the following command:
llama serve -hf ggml-org/MiniMax-Music3-GGUF
Note that when generating long songs exceeding about 3 minutes, it is recommended to specify a larger context length such as -c 20480 to ensure sufficient context.
How to Get and Use the Source Model (MiniMaxAI/MiniMax-Music3)
If you want to use the original source model MiniMaxAI/MiniMax-Music3, you can download it using the Hugging Face CLI or similar tools.
hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm
The source model supports multiple frameworks and can be deployed and used according to your needs in the following ways:
- SGLang-Omni: You can build an inference server via a shared speech API. It can be launched with the following command:
bash sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000In requests usingcurl, you pass the lyrics toinputand the music description toinstructionsto perform generation. - Diffusers: After installing from the pull request commit in development, you can call it from a Python script using
ModularPipeline. Various optimization features are available, such as automatic CPU offloading (enable_auto_cpu_offload) and group offloading per language model layer (apply_group_offloading). - ComfyUI: Official tutorials are provided, allowing you to integrate it into a node-based environment for music generation.
- Adding Prompt Expansion Skills: You can install the caption expansion tool with the following command:
bash npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter
Related Articles
- VibeVoice-ASR-GGUF Speech Recognition and Diarization Model: 8GB+ VRAM
- ggml-org Releases tinyLFM3-Omni-for-testing-gguf for CI
- ggml-org/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF: 12GB+ VRAM
- Qwen3-Coder-30B-A3B-Instruct-GGUF Text Generation Model: 24GB+ VRAM
What to Read Next
- Find models by VRAM (This model runs from the 12GB tier) → Other models that run on a 12GB GPU
- What Q8_0 mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- Learn about the publisher → MiniMax: models, licenses and articles
- Other models for the same task → Other audio models

