4-bit MLX conversions of open, permissively-licensed language models, used by FlashMe for on-device "bring-your-own-model" generation on refused content.
Each model is published as a Release whose assets are the MLX model directory (config.json, model.safetensors, tokenizer files, …). The app downloads them to a local directory and loads via loadModelContainer(directory:).
| Model | Source | License | Release tag |
|---|---|---|---|
| Salamandra 2B Instruct | BSC-LT/salamandra-2b-instruct | Apache-2.0 | salamandra-2b-instruct-4bit |
| EuroLLM 1.7B Instruct | utter-project/EuroLLM-1.7B-Instruct | Apache-2.0 | eurollm-1.7b-instruct-4bit |
All conversions: mlx_lm.convert -q (group_size 64, bits 4).