YuE2-3B-GGUF

Quantized GGUF weights of YuE2-3B and its audio decoder YuE2-Vae, prepared by LattifAI for local song generation in EdgeSpeak.

YuE2 turns lyrics into a full song with vocals and accompaniment. Before generating audio it can plan a score (melody, optionally with chords) that you can read, edit and feed back. All credit for the model goes to the YuE2 authors; this repository only changes the numeric format of the weights.

License: CC BY-NC 4.0, non-commercial use only. These files are quantized derivatives of the YuE2-3B and YuE2-Vae checkpoints and keep their original license. See License and attribution.

Files

File	Source	Format	Size	SHA-256
`YuE2-3B-Q5_K.gguf`	m-a-p/YuE2-3B	Mixed precision, Q5_K main	2,572,593,664 bytes	`1b13f9bd41e158e64c71659c01e119e367c6e2bff9fce0f50a2defa206a270f2`
`YuE2-VAEDecoder-Q6_K.gguf`	m-a-p/YuE2-Vae (decoder only)	Mixed precision, Q6_K main	59,699,264 bytes	`98b19a5d9a15a1296ca9003799ea06e3f2d005cccbb933e1eebbff38ed124583`

You need both files. The main model (3.58 billion parameters, 28 layers) generates the score and audio latents; the decoder turns latents into 48 kHz stereo audio.

YuE2-3B-Q5_K.gguf: 394 of 628 tensors are quantized with calibration; the remaining tensors keep their original precision.
YuE2-VAEDecoder-Q6_K.gguf: decoder weights only (the encoder is not included), quantized with calibration and a post-quantization refinement pass. Its source weights match model.safetensors in m-a-p/YuE2-Vae (SHA-256 807ce9d5149fa27c5ad3e6582058469852e908f6c5acc8c8aa338e7ab7751346).

Runtime

These files use a YuE2-specific tensor layout. Loading them with upstream llama.cpp has not been verified. They are loaded by the EdgeSpeak local engine and have been tested on CUDA (NVIDIA) and Metal (Apple Silicon).

Use with EdgeSpeak

In the EdgeSpeak desktop app, open Broadcast, choose YuE2, write lyrics with section tags such as [Verse] and [Chorus], and click Generate. You can pick how the score is planned, or paste your own ABC score.

The same model is available from the local OpenAI-compatible API:

curl --fail-with-body "$EDGESPEAK_BASE_URL/audio/speech" \
  -H "Authorization: Bearer $EDGESPEAK_API_KEY" \
  -H "Content-Type: application/json" \
  -H "buffered-speech: 1" \
  -d '{"model":"m-a-p/YuE2-3B","voice":"builtin:auto","input":"[Verse]\nStreetlights hum along the way\n[Chorus]\nSing it back to me","response_format":"wav","music":{"planning_mode":"full"},"output_container":"zip"}' \
  -o generation.zip

music.planning_mode is full (melody and chords, default), melody, or off. music.abc takes your own ABC score. With output_container: "zip" the response holds the audio, the score the model actually used, and metadata. See the API reference.

Quality

Measured against the original BF16 checkpoints on held-out data that was not used for calibration:

File	Held-out data	Result
`YuE2-3B-Q5_K.gguf`	54 complete records, 29,905 response positions	KL divergence 0.00285 (BF16 ∥ quantized), top-1 token agreement 95.0%; audio-latent cosine similarity 0.99968, MSE 0.000785
`YuE2-VAEDecoder-Q6_K.gguf`	6 songs, same fixed latents	SNR 32.55 dB, log-spectral distance 0.908 dB

These are model-level fidelity metrics. They do not include a word error rate for sung lyrics or a listening test, and they do not measure musical quality.

Speed

Measured in EdgeSpeak with both files on one 22.3-second song (48 kHz stereo):

Device	Generation time	Real-time factor
NVIDIA RTX 4090	9.09 s (median of 3 runs, model loading excluded)	0.41

Peak GPU memory sampled during these runs was about 5.6 GiB. Apple Silicon (Metal) runs these files but is currently much slower than real time; optimization is in progress. Speed varies with song length, device and settings, so treat these numbers as a single reference point.

License and attribution

The YuE2-3B and YuE2-Vae checkpoint weights are licensed by their authors under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). The files in this repository are adaptations of those weights and are distributed under the same license.

Original work: YuE2-3B (m-a-p/YuE2-3B) and YuE2-Vae (m-a-p/YuE2-Vae) by the Multimodal Art Projection (M-A-P) team.
Changes made: weights converted to GGUF and quantized (Q5_K / Q6_K mixed precision); the decoder half of YuE2-Vae extracted on its own. No retraining or fine-tuning.
Non-commercial: you may not use these files for commercial purposes. For commercial use, contact the YuE2 authors.
Upstream creator permission: the upstream MODEL_LICENSE adds a permission for individual creators to monetize outputs generated with YuE2. It does not cover commercial redistribution or sale of the weights. Whether it extends to these quantized files is determined by the upstream terms; read them before relying on it.
The upstream repositories carry separate notices and MIT license texts for the Oobleck / stable-audio-tools and SnakeBeta / BigVGAN code their inference implementation is derived from; see their THIRD_PARTY_NOTICES.md. This repository contains weights only, no code.

This repository is not affiliated with or endorsed by the YuE2 authors. Please cite their work:

YuE2 technical report: arXiv:2609.33757
SheetSage2: arXiv:2610.05336
Project: github.com/multimodal-art-projection/YuE

中文说明

YuE2-3B 及其音频解码器 YuE2-Vae 的 GGUF 量化权重，由 LattifAI 制作，用于 EdgeSpeak 在本地生成歌曲。

YuE2 能把歌词生成为带人声和伴奏的完整歌曲。生成音频前，它可以先规划一份乐谱（旋律，可附和弦），你可以查看、修改后再交给它使用。模型本身的功劳归 YuE2 作者；本仓库只改变了权重的数值格式。

许可证：CC BY-NC 4.0，仅限非商业用途。 这些文件是 YuE2-3B 与 YuE2-Vae 权重的量化衍生品，沿用原许可证，详见下文「许可与署名」。

文件

文件	来源	格式	大小	SHA-256
`YuE2-3B-Q5_K.gguf`	m-a-p/YuE2-3B	混合精度，主体 Q5_K	2,572,593,664 字节	`1b13f9bd41e158e64c71659c01e119e367c6e2bff9fce0f50a2defa206a270f2`
`YuE2-VAEDecoder-Q6_K.gguf`	m-a-p/YuE2-Vae（仅解码器）	混合精度，主体 Q6_K	59,699,264 字节	`98b19a5d9a15a1296ca9003799ea06e3f2d005cccbb933e1eebbff38ed124583`

两个文件都需要。主模型（35.8 亿参数，28 层）生成乐谱和音频 latent，解码器把 latent 还原为 48 kHz 双声道音频。

YuE2-3B-Q5_K.gguf：628 个张量中有 394 个经校准后量化，其余保留原精度。
YuE2-VAEDecoder-Q6_K.gguf：只含解码器权重（不含编码器），校准量化后另做了一轮量化后修正。源权重与 m-a-p/YuE2-Vae 的 model.safetensors 一致（SHA-256 807ce9d5149fa27c5ad3e6582058469852e908f6c5acc8c8aa338e7ab7751346）。

运行时

这两个文件使用 YuE2 专用的张量布局，尚未验证上游 llama.cpp 能否加载。目前由 EdgeSpeak 本地引擎加载，已在 CUDA（NVIDIA）和 Metal（Apple Silicon）上测试。

在 EdgeSpeak 中使用

在 EdgeSpeak 桌面版的播报页选择 YuE2，写好带 [Verse]、[Chorus] 等段落标签的歌词，点击生成即可。可以选择乐谱的规划方式，也可以粘贴自己的 ABC 乐谱。

本地 OpenAI 兼容 API 同样可用：

curl --fail-with-body "$EDGESPEAK_BASE_URL/audio/speech" \
  -H "Authorization: Bearer $EDGESPEAK_API_KEY" \
  -H "Content-Type: application/json" \
  -H "buffered-speech: 1" \
  -d '{"model":"m-a-p/YuE2-3B","voice":"builtin:auto","input":"[Verse]\n沿着街灯慢慢走\n[Chorus]\n唱给我听","response_format":"wav","music":{"planning_mode":"full"},"output_container":"zip"}' \
  -o generation.zip

music.planning_mode 取 full（旋律与和弦，默认）、melody 或 off；music.abc 传入自己的 ABC 乐谱；output_container: "zip" 时一次返回音频、模型实际采用的乐谱和元数据。详见 API 文档。

质量

与原始 BF16 权重在未参与校准的留出数据上对比：

文件	留出数据	结果
`YuE2-3B-Q5_K.gguf`	54 条完整记录，29,905 个回答位置	KL 散度 0.00285（BF16 ∥ 量化），top-1 token 一致率 95.0%；音频 latent 余弦相似度 0.99968，MSE 0.000785
`YuE2-VAEDecoder-Q6_K.gguf`	6 首歌，使用相同的固定 latent	SNR 32.55 dB，对数谱距离 0.908 dB

以上是模型层面的保真度指标，不包含歌词演唱的词错误率或听感测试，也不衡量音乐质量。

速度

在 EdgeSpeak 中用这两个文件生成同一首 22.3 秒的歌曲（48 kHz 双声道）：

设备	生成耗时	实时率
NVIDIA RTX 4090	9.09 秒（3 次中位数，不含模型加载）	0.41

这些运行中采样到的 GPU 显存峰值约 5.6 GiB。Apple Silicon（Metal）可以运行，但目前远慢于实时，仍在优化。速度随歌曲长度、设备和设置变化，以上数字只作为单个参考点。

许可与署名

YuE2-3B 与 YuE2-Vae 的权重由其作者以 CC BY-NC 4.0（署名-非商业性使用 4.0 国际）授权。本仓库中的文件是对这些权重的改编，以同一许可证发布。

原作品：Multimodal Art Projection（M-A-P）团队的 YuE2-3B（m-a-p/YuE2-3B）与 YuE2-Vae（m-a-p/YuE2-Vae）。
所做修改：权重转换为 GGUF 并量化（Q5_K / Q6_K 混合精度）；单独提取 YuE2-Vae 的解码器部分。没有重新训练或微调。
非商业：不得将这些文件用于商业目的。商业使用请联系 YuE2 作者。
上游创作者许可：上游 MODEL_LICENSE 另给个人创作者一项许可，允许将用 YuE2 生成的作品变现；该许可不包括权重的商业再分发或出售。它是否适用于本仓库的量化文件，以上游条款为准，使用前请自行阅读。
上游仓库为其推理实现所源自的 Oobleck / stable-audio-tools 与 SnakeBeta / BigVGAN 代码另附了声明和 MIT 许可证原文，见其 THIRD_PARTY_NOTICES.md。本仓库只含权重，不含代码。