YuE2-3B — MLX Serve (8-bit)

m-a-p/YuE2-3B with the YuE2-Vae decoder folded in and 8-bit affine (group size 64) on the backbone, in the layout of ahmadw/YuE2-3B-MLX (its 8bit/ folder, mirrored unchanged). Lyrics and a style prompt in, a 48 kHz stereo song out, planned from an editable ABC score.

Runs in mlx-serve (POST /v1/audio/music-generations and the app's Music tab):

mlx-serve pull ddalcu/YuE2-3B-MLX-Serve-8bit

Files: model.safetensors (backbone), vae.safetensors (decoder), config.json, vae_config.json, qwen.tiktoken, yue2_generation_config.json.

License

Weights are CC BY-NC 4.0 (non-commercial), inherited from YuE2-3B and YuE2-Vae. Credit YuE2 (m-a-p) and ahmadw's MLX conversion when you publish output. The SnakeBeta and Oobleck decoder code are MIT (NVIDIA, Stability AI); see the upstream THIRD_PARTY_NOTICES.md.

Downloads last month: 393

Safetensors

Model size

4B params

Tensor type

BF16

U32

MLX

Hardware compatibility

Quantized

Model tree for ddalcu/YuE2-3B-MLX-Serve-8bit

Base model

m-a-p/YuE2-3B

Quantized

(18)

this model