YuE2-3B-bf16 (MLX)

Non-commercial weights (CC BY-NC 4.0). Songs made with this model 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. Covers and supplied lyrics may need the original song's rights; this licence doesn't grant them.

The backbone of m-a-p/YuE2-3B (M-A-P; code multimodal-art-projection/YuE) for mlx-yue2-swift, a Swift-MLX port for Apple Silicon. YuE2 turns lyrics and a style prompt into a song with vocals: it plans a chord-annotated ABC score, then sings it (48 kHz stereo, up to 6 minutes). The score is editable and can be supplied back to re-render.

This repository is the 3.58 B Mixture-of-Transformers backbone. Its tensors are bf16 and bit-equal to the release. The VAE decoder is mlx-community/YuE2-Vae-fp32.

file	size
model.safetensors (628 tensors, bf16)	7.26 GB
qwen.tiktoken (tokenizer, as shipped by m-a-p)	2.6 MB
config.json (+ an mlx_yue2 section: dtype, source pin, generation defaults)	1.6 KB
LICENSE (HF, CC BY-NC 4.0), MODEL_LICENSE (GitHub, with the individual-creator permission), NOTICE, THIRD_PARTY_NOTICES.md, licenses/	—

Use (Swift, MLXEngine)

// Package.swift: .package(url: "https://github.com/xocialize/mlx-yue2-swift", from: "0.1.0") → product "MLXYuE2"
import MLXServeCore, MLXToolKit, MLXYuE2

let engine = MLXServeEngine()   // CC-BY-NC-4.0 is outside the permissive policy: it registers with a licence advisory
let id = try await engine.register(YuE2Package.registration,
                                   configuration: YuE2Configuration(tier: .bf16, modelsRootDirectory: modelsRoot),
                                   id: "yue2-3b")
try await engine.prepare(.music, package: id)   // downloads this repo and YuE2-Vae-fp32 on first use
let song = try await engine.run(MusicRequest(caption: "English, warm piano pop, expressive female voice, 88 BPM",
                                             lyrics: "[Verse]\n…\n\n[Chorus]\n…", seed: 831001), package: id) as! MusicResponse
// song.audio: 48 kHz stereo WAV; song.score: the ABC plan (edit it, pass it back as MusicRequest.score)

Tiers are quantised at load from this bf16 repo: .bf16, .ar8 (the AR path 8-bit), .int8 (the AR and NAR paths 8-bit). Memory (phys_footprint at the 6-minute worst case, M5 Max): bf16 7.6 GB resident / 14.6 GB peak, ar8 6.4 / 13.3, int8 5.1 / 11.9. An English 58 s song renders in 41 s at bf16 on an M5 Max. Planning modes full and off; melody is not offered.

Parity

The Swift port is gated against an fp32 PyTorch oracle of the release: tokenizer id-exact; AR logits argmax 100 % (relL2 ≤ 1.8e-6); NAR velocity ≤ 1.6e-5; VAE ≤ 4.6e-6; end-to-end audio ≤ 1.7e-4. bf16 and 8-bit tiers sit 1.2–1.6e-2 from fp32 in velocity, where the release itself runs. Details: MEASUREMENTS.md.

Attribution

YuE2, YuE2-3B, https://huggingface.co/m-a-p/YuE2-3B, by M-A-P, under CC BY-NC 4.0. The weights here are unmodified apart from serialisation; NOTICE lists exactly what changed. Not endorsed by M-A-P.