Apple ML Research: Muse Spark Open with MLX on Mac Studio
Points and comments are a snapshot, not live.
Apple shows Muse Spark Open across four Mac Studios via RDMA and MLX.
Apple demonstrates distributed LLM inference with MLX and RDMA over Thunderbolt on Mac Studio clusters. A single M5 Ultra runs Qwen 3.8 and Mistral Medium 3.5. Muse Spark Open (816B parameters, NVFP4) requires multiple machines. With four M5 Ultras using tensor parallelization, it achieves 36.73 tokens/s decode speed. Pipeline parallelization speeds prefill by overlapping chunks. Power consumption is under 300W per machine.
What commenters are saying
Commenters note the cost. A 512GB Mac Studio is likely over $9,499; four could exceed $40,000. Some find API costs more practical, while others note local inference may suit contractual or legal requirements. There is surprise Apple featured a Meta model prominently. A few criticize the benchmarks, saying Qwen is handicapped by being tested at BF16 while Muse uses NVFP4. The trend toward smaller, moreefficient models is seen as favoring Apple's long-term position.