majentik

/

gemma-4-26B-A4B-it-TurboQuant-MLX-8bit

like 0

LM Studio’s MLX backend ships with full Gemma 4 support already bundled — mlx-vlm 0.4.3, mlx-lm 0.31.2, and the complete gemma4 model module. However, a manual guard in generate.py raises a ValueError before the model ever loads:

There is a solution

Nice, thanks for that. Found my way

Pretty cool model, overall less memory usage (55GB vs. 40GB) but it has large spikes once it begins running a prompt then comes back down to the baseline, and sometimes even to down to as if it wasn't in memory.