mlx_vlm.convert --hf-path INSAIT-Institute/BgGPT-Gemma-3-4B-IT \
  --mlx-path BgGPT-Gemma-3-4B-IT-MLX-4bit \
  --quantize --q-bits 4 --trust-remote-code

The converter reported 5.475 bits per weight.

Local Smoke Test

Verified on an Apple Silicon Mac through mlx-openai-server using the text-only lm handler:

bash

mlx-openai-server launch \
  --model-path ./BgGPT-Gemma-3-4B-IT-MLX-4bit \
  --model-type lm \
  --served-model-name bggpt-gemma-3-4b-it-mlx-4bit \
  --host 127.0.0.1 --port 8081 \
  --disable-batching

Prompt:

Отговори кратко на български: Какво е автокъща?

Observed response:

Автокъща е търговски обект или фирма, която продава автомобили – нови или употребявани.

Notes

This is an MLX conversion of the original INSAIT checkpoint, not a new finetune.
Text chat is verified. Multimodal serving was not validated in this upload.
The original model is distributed under the Gemma Terms of Use; this conversion preserves that license metadata.

flexi767/BgGPT-Gemma-3-4B-IT-MLX-4bit

BgGPT-Gemma-3-4B-IT-MLX-4bit

Local Smoke Test

Notes

flexi767/
BgGPT-Gemma-3-4B-IT-MLX-4bit