ckadirt/MiniMax-Music3-GGUF

MiniMax Music 3 GGUF

Experimental artifact snapshot, not an approved production release.

The source validation matrix records dense parity, quantized profiles, CPU/CUDA real-model smoke tests, and pause/resume checks as passed. Vulkan, Metal, and Android real-model validation remain pending; release_approved is false. These are recorded upstream results, not validation runs performed for this upload.

Portable GGUF components for MiniMaxAI/MiniMax-Music3, converted for minimax-music3.cpp and the Cantor engine ABI v1.

These files are checkpoint derivatives. They are governed by the included MODEL_LICENSE (the MiniMax Music 3 Community License), not by the Apache-2.0 license used for the C++ source repository. Review the model license before using or redistributing them.

Components

Load exactly one file for each role: lm, rvq, condition, dit, and vae. The condition encoder has one F32 artifact. The VAE has F32 and F16 storage variants. LM, RVQ, and DiT provide the following quality ladder:

• BF16 for LM/RVQ or F32 for DiT: reference storage
• Q8_0: high-quality quantized storage
• Q6_K: quality-oriented K-quant
• Q5_K_M: intermediate K-quant
• Q4_K_M: balanced/default K-quant

The recommended balanced set is lm-Q4_K_M, rvq-Q4_K_M, condition-F32, dit-Q4_K_M, and vae-F16. Do not rename component files: their exact bytes are part of Cantor pause/resume identity checks.

Artifact	Bytes	SHA-256
lm-BF16.gguf	17,180,559,616	17e8f57e5712ae47f3b6882fe9756b00fb648d6d8376029fbf1ab02daff5dbe8
lm-Q8_0.gguf	9,132,902,656	a536342228fdb88f156d4caeb767f03c2cf4eb788815c71d4065c0e1330e3e23
lm-Q6_K.gguf	7,053,924,608	96a8228332c7d8897ee560d524a596ba9ff30346c9821db47b0b974d675170bb
lm-Q5_K_M.gguf	6,452,304,160	d4e061ae16a0cb664964749b5f603755a0fe26c60d81a3bb66e95837b7ac28d6
lm-Q4_K_M.gguf	5,886,073,120	5f67c61c1c130d2d8f0c8f30098eb1c9396c8d46ac9bbce2e212fc712340925b
rvq-BF16.gguf	1,292,063,296	10f34b7e4c3e6e49866ae39ca70bb72ec579654d80d43d3c58f72680035b3dbd
rvq-Q8_0.gguf	686,641,728	2ab61b908e9509d9484f1ff7c86278788a32e670a985e1e27d15b18fb753faf3
rvq-Q6_K.gguf	530,207,296	c88519ff7473f08fa4c03ed087db49d011531b44a2ca2e9aa23c9ec5e8049a03
rvq-Q5_K_M.gguf	474,501,696	ad99404155fe8523dce2b4543d1e73a708040f40e8004f73477d97b6b5298072
rvq-Q4_K_M.gguf	422,072,896	f11c2408e4c7e6a606e126ffbd43b29db78dece6b9f3b5428436fcd9624d6477
condition-F32.gguf	100,679,296	a1a89b8e9ec9ac3311fecfd8fe59d5fc4cf0b7776e09cc53e90423a1014fbd6d
dit-F32.gguf	9,727,688,128	f417afed1906598322e72aaf8353c0aa17a90a43d51abc135500e5184fb6fbe1
dit-Q8_0.gguf	2,616,286,688	c39e6f73a0eb18518ad89beb9fe8ca71dd323429d69e229300061af4e7e8032d
dit-Q6_K.gguf	2,029,975,008	d875ec9ce0d6ab0bbef0f97da8dbe66e7f2ce627190c10cf927eb0ad7477ae65
dit-Q5_K_M.gguf	1,808,719,328	e60d052635f32bfe0f3a89fd7e7f85c14edd9720d8c9502733d9b89456de7bd9
dit-Q4_K_M.gguf	1,600,478,688	8cbaeb770f66804178695846f096b1b21a982f4425d368ca2c2c6bba6455729d
vae-F32.gguf	216,706,336	4ec9da855b543e394c8abf20ba43c6a84d5b3d1f20cc9a22467195221cf68df9
vae-F16.gguf	108,498,016	68bdb356f42168708063567c21ec0764e8fd78b8726d863ea70a11b855d359cd

Each GGUF has a .manifest.json and .sha256 sidecar. The LM GGUFs also embed the pinned tokenizer, all component configs, source inventory, model license, and model specification. Every GGUF embeds its component, profile, source commit, source-inventory hash, converter version, and GGML revision.

Runtime

Native output is stereo 44.1 kHz; Cantor requests may explicitly select the Torchaudio-compatible 32 kHz serving resampler. The same model implementation targets CPU, CUDA, HIP, Vulkan, Metal, and Android through GGML.

This snapshot preserves the existing validation status. See validation/release-matrix.json for the recorded gates.