MiniMax-Music3-GGUF

MiniMax Music 3 is an open-weights text-to-music generation model from MiniMax AI that creates complete, structurally coherent songs up to five minutes long in 32 kHz, 16-bit stereo WAV format. Conditioned on section-tagged lyrics (such as [Verse] and [Chorus]) and detailed style descriptions covering metadata, vocals, and arrangement, it uses a hierarchical Hybrid-LM architecture pairing a Qwen3-initialized 8B Global LLM for long-range musical structure with a 0.6B Local LLM for frame-level acoustic detail. Instead of decoding directly from discrete Residual Vector Quantization (RVQ) tokens, the system fuses the hidden states of both language models and synthesizes high-fidelity audio through a 2.4B Flow Matching module and Flow-VAE decoder, supporting deployment via SGLang-Omni, Hugging Face diffusers, and ComfyUIβ€”even scaling down to 8 GB VRAM GPUs with layer offloading.

Model Files

File Name Quant Type File Size File Link Description
MiniMax-Music3.BF16.gguf BF16 17.2 GB Link Full BF16 weights. Highest quality, largest file size.
MiniMax-Music3.Q3_K_L.gguf Q3_K_L 4.68 GB Link Lower quality but usable, good for low RAM availability.
MiniMax-Music3.Q3_K_M.gguf Q3_K_M 4.37 GB Link Low quality, smallest file size.
MiniMax-Music3.Q4_K_M.gguf Q4_K_M 5.3 GB Link Good quality, default size for most use cases, recommended.
MiniMax-Music3.Q4_K_S.gguf Q4_K_S 5.08 GB Link Slightly lower quality with more space savings, recommended.
MiniMax-Music3.Q5_K_M.gguf Q5_K_M 6.15 GB Link High quality, recommended.
MiniMax-Music3.Q5_K_S.gguf Q5_K_S 6.02 GB Link High quality with slight space savings, recommended.
MiniMax-Music3.Q6_K.gguf Q6_K 7.05 GB Link Very high quality, near perfect, recommended.
MiniMax-Music3.Q8_0.gguf Q8_0 9.13 GB Link Extremely high quality, generally unneeded but max available standard quant.
MiniMax-Music3.mmproj-bf16.gguf mmproj-bf16 6.64 GB Link Multimodal projection file in BF16 format. Used for audio/music capabilities.

Run with llama.cpp

Install dependencies and clone llama.cpp

!pip install transformers huggingface-hub

!git clone https://github.com/ggml-org/llama.cpp.git
%cd /data/llama.cpp

Download GGUF model files

Download the BF16 model and BF16 multimodal projector:

from huggingface_hub import snapshot_download

# Target local directory
local_dir = "./model/MiniMax-Music3-GGUF"

# Download only the specified files
snapshot_download(
    repo_id="prithivMLmods/MiniMax-Music3-GGUF",
    local_dir=local_dir,
    allow_patterns=[
        "MiniMax-Music3.BF16.gguf",
        "MiniMax-Music3.mmproj-bf16.gguf",
    ],
    resume_download=True,
)

Build llama-tts

Build the TTS CLI with the CUDA backend enabled:

!cmake --build . --config Release -j$(nproc) --target llama-tts

Music generation

Define the music generation prompt:

prompt = """[Tags]
Genre: Uplifting cinematic pop with acoustic guitar, warm piano, ambient synths, and energetic drums.
BPM: 108.
Mood: Hopeful, emotional, dreamy, and uplifting.
Vocals: Expressive male lead vocals with clear melodic delivery and layered harmonies in the chorus.
Production: Polished studio recording with spacious reverb, a rich instrumental arrangement, and a powerful melodic chorus.

[Lyrics]
[intro]
Oh, tonight we come alive
Underneath the endless sky

[verse]
Under the silver moonlight, we chase the stars tonight
Every little dream we hold is shining twice as bright
Through the endless open sky, we'll never say goodbye
With every heartbeat racing, we feel that we can fly

[pre-chorus]
Take my hand and don't let go
There's a thousand lights below

[chorus]
We are young, we are free
This is where we're meant to be
Let the music take us higher
Burning like a midnight fire
Through the night, through the rain
We will never fade away
We are young, we are free
Forever chasing destiny

[bridge]
When the world is falling silent
And the stars begin to fade
We'll remember every moment
And the promises we made

[final chorus]
We are young, we are free
This is where we're meant to be
Let the music take us higher
Burning like a midnight fire

[outro]
Underneath the silver sky
We'll keep this feeling alive"""

Music generation (GPU offloading)

Run inference using llama-tts with GPU offloading enabled:

! /data/llama.cpp/build/bin/llama-tts \
    -m /data/llama.cpp/model/MiniMax-Music3-GGUF/MiniMax-Music3.BF16.gguf \
    -mm /data/llama.cpp/model/MiniMax-Music3-GGUF/MiniMax-Music3.mmproj-bf16.gguf \
    -p "$prompt" \
    -ngl 99 \
    -n 4096 \
    --seed 42 \
    --output /data/llama.cpp/model/MiniMax-Music3-GGUF/out.wav

Output

The generated audio file will be saved at:

/data/llama.cpp/model/MiniMax-Music3-GGUF/out.wav

llama.cpp

LLM inference in C/C++ β€” https://github.com/ggml-org/llama.cpp

Quantized using llama.cpp by prithivMLmods.

Downloads last month
321
GGUF
Model size
9B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for prithivMLmods/MiniMax-Music3-GGUF

Quantized
(29)
this model

Collections including prithivMLmods/MiniMax-Music3-GGUF