mlx-community

/

medgemma-27b-it-bf16

like 4

Follow

MLX Community 11.8k

Hi all — I’m new here 👋
I’d like to test MedGemma 27B multimodal locally on Apple Mac Studio and size the hardware correctly. Could you share real-world numbers and tips?

My key questions

Unified memory:
Do I really need ≥160 GB shared memory for comfortable inference? (fine for me, if I do need it) using Mac Studio M3 Ultra (quant 16 bit)

Latency per image (MRI):
End-to-end time for a single 1024×1024 MRI slice (PNG/JPEG), batch=1, simple radiology prompt.
I will use M3 ultras Apple mac Studio. MLX Version will speed it up significantly, right?

Image I/O & preprocessing:
For medical images, did you convert DICOM → PNG/JPEG first? Any pitfalls with 16-bit grayscale or windowing?

Thanks in advance
Axel