Apple releases MLX update to run 816-billion-parameter model across Macs

Apple released an MLX update that it says runs the 816-billion-parameter Muse Spark Open model across four Mac Studios at over 36 tokens per second. The open-source framework distributes inference across Apple silicon computers, allowing developers to serve the model locally for tools such as coding agents. The setup uses low-latency communication over Thunderbolt and splits each model layer across machines. Apple tested the model in NVFP4, a four-bit format that reduces its memory footprint to about 493GB. Intermediate computations still prevent it from fitting on one Mac Studio. At a 65,536-token context, communication overhead made two-machine tensor parallelization slower at processing input than splitting successive layers between machines.