A smaller copy of VisualInference/MOSS-Image-SFT-MLX-compact for Macs with 8 GB of memory, made for and used by Protoart Studio. Runs with moss-image-mlx.
What was changed (Apache-2.0 §4b): only the picture model (transformer/). It was re-stored from 8-bit to
4-bit (group 64) except attention (q, k, v, o), the text embedder and the first and last blocks, which stay 8-bit
(~5.3 GB → 3.71 GB). The layers kept in bf16 by the original are unchanged. The text encoder, tokenizer, scheduler and
VAE decoder are the compact tier's files, unchanged. Quantization spec: in transformer/config.json.
Expect slightly less fine detail than the 8-bit original; lettering is decided by the unchanged text encoder and the 8-bit attention.
See LICENSE and NOTICE. Not affiliated with OpenMOSS.
text_encoder/model-00001-of-00002.safetensors 4.3 GBtransformer/model.safetensors 4.0 GBtext_encoder/model-00002-of-00002.safetensors 2.2 GBvae/decoder.safetensors 112 MB