MiniMax Music 3 WebGPU artifacts

Converted browser artifacts for MiniMax Music 3 WebGPU, a local music generation demo that runs the official MiniMaxAI/MiniMax-Music3 checkpoint on WebGPU in your browser.

The demo is available at https://li-na.github.io/minimax-music3-webgpu/. After the one-time model download, every generation happens entirely inside the browser. Prompts, lyrics, and generated audio are never sent to any server.

MiniMax Music 3 WebGPU is a community project unaffiliated with MiniMax and is neither sponsored nor endorsed by them.

Contents

This repository holds the music-variable release. manifest.json at the repository root lists every artifact with its byte length and SHA-256; the application reads it first and resolves every other path relative to it, so the layout has to stay intact.

Path Bytes Files Stage
global/ 5,694,616,196 54 Semantic decoder, reduced head, embedding shards, tokenizer
flow/ 1,377,818,234 12 Flow-matching transformer
rvq/ 1,292,007,411 14 RVQ depth decoder, feedback network, embedding
vocoder/ 108,383,113 2 Vocoder
condition/ 50,336,945 2 Condition encoder
manifest.json 36,224 1 Release descriptor

Total: 8,523,198,123 bytes across 85 files. Every artifact is verified against its manifest SHA-256 after download, so the CDN does not have to be trusted.

Conversion

Item Value
Source checkpoint MiniMaxAI/MiniMax-Music3 at fbdf52f
diffusers 3681e65
Weight quantization 4-bit block-wise, block size 32, symmetric, accuracy level 4
Flow exceptions time_proj.weight and proj_out.weight stay FP16 MatMul weights
Convolutions float16
Snake activations blocks.0.snake1 and blocks.1.snake1 in FP32
Runtime ONNX Runtime Web on the WebGPU execution provider

The conversion pipeline lives in the application repository under tools/converter.

Requirements

Item Requirement
Browser Desktop Chromium (Chrome 151 or later) with WebGPU and shader-f16 support
GPU memory 12 GB of VRAM for workloads up to one minute, 16 GB for the full capacity
Storage About 8 GiB of browser storage for the model cache

The VRAM guidance was measured on the preserved block-size-128 release and is for reference only. Memory and longer-duration timing measurements for this release are still pending; its five-second generation gate passes.

Usage

Open the hosted demo, or point your own build at this repository by pinning the manifest to a commit:

VITE_MINIMAX_MANIFEST_URL=https://huggingface.co/hyung778/minimax-music3-webgpu/resolve/<commit-sha>/manifest.json

Pin a commit rather than main. A build that follows a branch would pick up a different set of weights the moment this repository is updated.

License

The MiniMax Music 3 model is distributed as open weights under the MiniMax-Music3 Community License, which is not an OSI open-source license. These converted artifacts retain that license and its notices. Use of the artifacts and of generated audio is subject to the license terms, including its Acceptable Use Policy, and generated audio distributed publicly must be clearly disclosed as AI-generated.

MiniMax Music 3 and related assets are the property of MiniMax.

The application code is licensed separately under MIT.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for hyung778/minimax-music3-webgpu

Quantized
(22)
this model