YuE2-Vae-fp32 (MLX)

Non-commercial weights (CC BY-NC 4.0). Songs made with YuE2 may not be sold, licensed or used in paid work, except under M-A-P's individual-creator permission (MODEL_LICENSE, 2026-09-16): personal users, content creators and musicians acting in an individual capacity may publish, sell and license songs they make with it. Companies may not use the model commercially without a licence from M-A-P (ryuanab@connect.ust.hk), and a hosted service making songs for paying users is commercial use. Responsible use is a condition of the permission.

The listening VAE decoder of m-a-p/YuE2-Vae (M-A-P), converted for MLX. It is used by mlx-yue2-swift, the Swift-MLX port of YuE2-3B, together with mlx-community/YuE2-3B-bf16 (usage on that card). It is an Oobleck decoder: 64-dim latents at 25 Hz β†’ 48 kHz stereo, fp32, decoded in exact tiles.

file size
decoder.safetensors (173 tensors, fp32; weight_norm folded, convs in MLX layout, encoder dropped) 265 MB
config.json 1.4 KB
LICENSE (HF, CC BY-NC 4.0), MODEL_LICENSE (GitHub, with the individual-creator permission), NOTICE, THIRD_PARTY_NOTICES.md, licenses/ (Oobleck / SnakeBeta MIT) β€”

Parity: the folded tensors equal upstream's own remove_weight_norm output (max |Ξ”| 0). The Swift decode matches the fp32 PyTorch oracle to relL2 ≀ 4.6e-6, and the tiled decode equals the full decode within 2.7e-6.

Attribution

YuE2, YuE2-Vae, https://huggingface.co/m-a-p/YuE2-Vae, by M-A-P, under CC BY-NC 4.0. NOTICE lists exactly what changed (the decoder only, weight_norm folded, layouts transposed). Not endorsed by M-A-P.

Downloads last month
-
MLX
Hardware compatibility
Log In to add your hardware

Quantized

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

Model tree for mlx-community/YuE2-Vae-fp32

Base model

m-a-p/YuE2-Vae
Finetuned
(2)
this model