MiniMax-Music3 11B INT8: full songs from a description and lyrics
This is a single, self-contained checkpoint set for MiniMax-Music3. From a text
description of the music and lyrics tagged by section ([verse], [chorus], …), it
generates a complete song with vocals, up to five minutes long by design, as 44.1 kHz stereo.
The bundle pairs the INT8 + ConvRot pruned text encoder (the autoregressive music model) with the INT8 + ConvRot DiT and the DAV decoder. That is ~11.9 GB, instead of the ~57 GB upstream release. The tokenizer is stored inside the text-encoder file, so everything loads locally from this directory and nothing is downloaded at runtime.
Size. MiniMax does not publish a parameter count. The "11B" in this repository's name is ours, counted from these checkpoints (≈ 10.9 B parameters):
| Part | Parameters |
|---|---|
| Autoregressive music model (Qwen3-8B fine-tuned by MiniMax), transformer blocks | 6.95 B |
| Its embeddings and c0 head, after pruning | 0.79 B |
| RVQ depth decoder | 0.62 B |
| DiT | 2.46 B |
| DAV decoder | 0.05 B |
Precision. The text encoder and the DiT carry INT8 + ConvRot weights. The embeddings and the c0 head stay bf16, and the small DAV decoder is fp32.
Credits
| Original model | MiniMax: MiniMax-Music3. A Qwen3-8B-based autoregressive music model with an RVQ depth decoder, a flow-matching DiT, and the DAV audio decoder |
| Pruning, quantization and repack | Comfy-Org: the pruned INT8+ConvRot text encoder, the INT8+ConvRot DiT and the DAV file in this repository are their single-file repack of MiniMax-Music3, prepared for ComfyUI |
deAPI-ai did not train or quantize any of these weights. This repository selects three files
from Comfy-Org's repack, byte-identical to them (SHA-256 in MANIFEST.json), and
collects them into one directory together with the original license.
Contents
| Path | Component | Size | Format |
|---|---|---|---|
text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors |
The "text encoder": tokenizer + the Qwen3-8B-based autoregressive model + RVQ depth decoder that turns caption and lyrics into the song's acoustic codes | 9.2 GB | pruned (prompt embedding and LM head cut to the usable rows), INT8 + ConvRot; the tokenizer is embedded as a tensor |
diffusion_models/minimax_music3_dit_int8_convrot.safetensors |
Flow-matching DiT that renders those codes into audio latents | 2.5 GB | INT8 + ConvRot |
vae/minimax_music3_dav.safetensors |
DAV decoder, latents → 44.1 kHz stereo waveform | 217 MB | fp32 |
LICENSE |
MiniMax-Music3 Community License, unchanged from the original repository | 7 KB | — |
Directory naming. The folders follow Comfy-Org's layout (text_encoders, plural).
The original MiniMaxAI release is a diffusers-style multi-folder layout, and it has no
counterpart for these single-file checkpoints.
How it generates
- Autoregressive stage. The text encoder reads the caption and lyrics. It then samples the song 25 frames per second: one coarse code from the language model, then 7 finer codes from the RVQ depth decoder. Classifier-free guidance is applied at every code. The song ends when the model emits its end token, typically once the lyrics' section tags run out. The requested duration is only an upper limit.
- Diffusion stage. The DiT denoises audio latents conditioned on that sequence: 30 Euler steps, CFG 1.7, in overlapping windows.
- Decoding. The DAV turns the latents into a 44.1 kHz stereo waveform.
Notes on the components
- Prompting. The caption is a free-text description of the music. The model card's
recommended structure is Global Metadata → Vocal Details → Arrangement: genre and tempo,
vocal gender and timbre, and an arrangement described as a timeline. Lyrics carry one
section tag per line (
[intro],[verse],[pre-chorus],[chorus],[bridge],[instrumental],[solo],[outro]), with the lines of that section below it. A single[Instrumental]tag yields a piece of only a few seconds; for a longer instrumental, write out several sections. See MiniMax's prompting guide. - Sampling defaults (the values of the ComfyUI reference workflow). DiT: 30 steps, CFG 1.7. Autoregressive model: CFG 1.7, top-k 50.
- No conditioning from audio. MiniMax did not release the RVQ or VAE encoders, so reference audio, continuation and inpainting are not possible with these weights. Tempo, key and meter are expressed in the caption, not as separate controls.
- INT8 + ConvRot stores the weights in a rotated basis (group size 256), and the matching kernels apply the rotation to the activations. It is the same format Comfy-Org uses for its other INT8 repacks.
- Footprint. ~11.9 GB of weights. With the stages swapped through pinned host memory, generation peaks around 8–10 GiB VRAM for songs up to 3 minutes, and the host needs ≥ 32 GB of system RAM.
Licenses
| Component | Source | License |
|---|---|---|
| Text encoder, DiT, DAV (original weights) | MiniMaxAI/MiniMax-Music3 | MiniMax-Music3 Community License |
| Pruned / INT8+ConvRot single-file repack of the above | Comfy-Org/MiniMax-Music-3 | MiniMax-Music3 Community License (derivative of the above) |
| Language-model base | Qwen/Qwen3-8B, fine-tuned by MiniMax | Apache-2.0 upstream; redistributed here under the MiniMax-Music3 license |
| DiT code lineage | Stability-AI/stable-audio-tools | MIT |
| DAV code lineage | descriptinc/descript-audio-codec | MIT |
NOTICE
MiniMax-Music3 COMMUNITY LICENSE — Copyright (c) 2026 MiniMax. The full license, including its Acceptable Use Policy, is in
LICENSE.
Modified files. The text-encoder and DiT weights in this repository are not the original files. They are pruned and quantized derivatives (INT8+ConvRot) produced by Comfy-Org, and the DAV is their format conversion of the original decoder. deAPI-ai redistributes them unchanged.
Before you use these weights, read the license. It carries an acceptable-use policy and has further conditions:
- a commercial product or service that uses the model must display "MiniMax-Music3" prominently in its user interface;
- products or services generating over 20 million USD in yearly revenue need separate written authorization from MiniMax;
- anyone offering generation from the model to third parties must maintain safeguards against uses and outputs that violate the license, including third-party intellectual-property rights.
The bundle is made available by deAPI-ai on Hugging Face, configured for integration with GamerHash AI's platform. All credit for the underlying models belongs to their respective creators listed above.
Hosted by deAPI-ai for use with GamerHash AI