Apple runs Meta's full 816B Muse Spark Open locally on two M5 Ultra Mac Studios via MLX

Apple's collaboration with Meta turns the 816B Muse Spark Open into a desktop story. Two 512GB M5 Ultra Mac Studios, paired and running MLX, sustain more than 36 tokens per second. That speed is practical for agent loops rather than just a demo.

The setup works because of unified memory. The 42B-active MoE architecture keeps per-token compute modest, and roughly 1TB of pooled memory holds all the weights.

The timing sets up a direct contest with NVIDIA, whose RTX Spark laptops (128GB unified memory) shipped the same day. Even with this demo, the price tag of two maxed-out Mac Studios keeps the setup in prosumer territory.