YuE2-3B (NF4, saved)

Derived from a copy of m-a-p/YuE2-3B at revision 8f24312f187763b9854adeec87e437e2b03bbff1, by m-a-p. This is not an official release.

Changes

The transformer blocks quantised to 4 bit (bitsandbytes NF4, bf16 compute) and saved; the output layer (lm_head) and the parts the NAR step adds (llm2vae, vae2llm, time_embedder, latent_pos_embed) stay in bfloat16. No retraining.

Licence

Creative Commons Attribution-NonCommercial 4.0, as the source — see LICENSE.

Non-commercial use only, with attribution, as the licence requires. The licence text in LICENSE includes the licensors' additional permission for individual creators.

All rights to the model remain with its authors. SHA256SUMS, where present, lists the checksum of every file.

The file ok (one line) marks the folder as complete for Audante's YuE2 engine (engines/yue2_run.py): it loads these weights as they are and quantises nothing.

Downloads last month: 58

Safetensors

Model size

4B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Audante/YuE2-3B-nf4

Base model

m-a-p/YuE2-3B

Quantized

(25)

this model