OpenAI's new 120B MoE runs nicely in mlx-lm on an M3 Ultra. Running the 8-bit quant:
38,44K