VincentGOURBIN/yue2-mlx-packs

YuE2 MLX packs — poids pré-quantifiés pour yue2-mlx-swift

Dérivés de m-a-p/YuE2-3B (CC-BY-NC-4.0) pour https://github.com/VincentGourbin/yue2-mlx-swift (MIT). Usage non commercial, attribution m-a-p.

These packs power PocketAnthem on the App Store (https://apps.apple.com/us/app/pocketanthem/id6815755463): lyrics and style → a song generated entirely on the iPhone, int4-mixed-head with the 4bit-lean reference configuration. Site: https://vinceforge.com

Pack	Bits	Size	Profiles
int4-mixed-head	AR / embeddings / head 4, NAR 8	2.5 GB	4bit-fast, 4bit-lean, 4bit-tiny
qint8-all-head	8 throughout	3.5 GB	8bit-fast, 8bit-lean
int4-head	AR / embeddings / head 4, NAR bf16	4.4 GB	the fastest on a Mac

Download with the engine: yue2 download --model int4-mixed-head (SHA-256 verified). Format safetensors MLX (poids packés uint32 + scales + biases, groupe 64, affine), métadonnées format=yue2-prequantized-v1. Voir docs/Weights.md du dépôt.

Support files (support/)

The small files the engine reads next to a pack, so a pack downloaded into an empty models directory loads on its own (yue2 download --model <pack> fetches them into YuE2-3B/, verified against support/SHA256SUMS):

• config.json, yue2_generation_config.json — copied unchanged from m-a-p/YuE2-3B (CC-BY-NC-4.0);
• tokenizer.json, tokenizer_config.json — the tokenizer of Qwen/Qwen2.5-0.5B (Apache-2.0), converted by Scripts/convert-tokenizer.py (special-token recognition removed so it matches YuE2's encode_ordinary; token-for-token identical to the checkpoint's qwen.tiktoken).