Geode 路 YuE2-3B for phones

YuE2-3B and YuE2-Vae by m-a-p, converted for Geode, an Android app that runs models on the phone itself: the song's score, its singing and its band are made on the device, nothing is sent anywhere.

The app downloads one of these packs from its Music settings:

Pack Size For
phone/ 5.45 GiB every phone: the processor computes the sound (slow without an NPU matrix unit: start with short songs)
phone-npu/ 8.08 GiB Snapdragon 8 Gen 2 and newer: the same, plus the AR's fp16 layers for the NPU (the sound's prompt on HMX)

What is inside

  • embed.bin - the token embeddings (bf16).
  • ar_q8.bin - the 3B transformer for the generation (the ABC score and the codec tokens): int8 weights, per-row scales.
  • nar_f16.bin - the acoustic stage (flow matching, 32 midpoint steps) in fp16, laid out for the Hexagon NPU's matrix unit; the processor reads the same file.
  • ar_f16.bin (phone-npu/ only) - the transformer's layers in fp16 for the NPU: the acoustic stage's prompt.
  • vae.bin - YuE2-Vae's decoder (25 Hz latents to 48 kHz stereo).
  • tokenizer.bin - the model's tokenizer.
  • geode.json - the pack's manifest (files, tensors, configuration).

The settings are the reference's: cot "full", 32 midpoint ODE steps, context 24576, CFG per song, mastering to -12.5 LUFS with at most +4 dB. Measured against the reference's fp32 run, the phone's acoustic latents differ by 1.2e-3 (relative L2) - less than the reference's own bf16 run.

Licence

The weights are CC BY-NC 4.0, as the originals: non-commercial use only, with attribution to the YuE2 authors. See LICENSE and the original models' third-party notices.

YuE2 by the Multimodal Art Projection team: https://github.com/multimodal-art-projection/YuE

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for Maximilian228/Geode-YuE2

Base model

m-a-p/YuE2-3B
Finetuned
(12)
this model