Apple and Meta run Muse Spark Open on Mac Studio clusters

Apple's machine learning team has released an updated iteration of its open-source MLX framework, designed to leverage Apple silicon for high-performance local AI inferencing. Developed alongside Meta, the latest software additions optimize distributed execution for Meta's open-weight Muse Spark Open model across interconnected hardware. The updated framework introduces software support for Neural Accelerators inside M5 family GPUs, low-latency Remote Direct Memory Access over Thunderbolt, and improved implementations for mixture of experts architectures.

By taking advantage of Apple's unified memory architecture, hardware setups can handle near-trillion-parameter frontier models on-premises without cloud infrastructure. Meta noted that an M5 Ultra chip lets integrated GPUs access up to 512 GB of memory per Mac Studio at 1.2 TB/s bandwidth. Linking two Mac Studios allows them to act as a unified supercomputer with a terabyte of memory to host an entire model. In benchmarks featuring a four-unit Mac Studio cluster, Muse Spark Open generated over 36 tokens per second, surpassing typical human reading speeds.

For AI practitioners and enterprise developers, this capability provides a cost-effective, private alternative to hosting large language models in external data centers or paying recurring cloud inference fees. Meta demonstrated the real-world utility of these Mac clusters by using them to execute agentic workflows entirely on-device, including writing bug fixes and applying new features to an open-source Mastodon client and driving DaVinci Resolve to review video footage and execute tasks. As executives like Apple CEO John Ternus noted, the approach lets technical teams run massive agentic models directly from their desks while maintaining strict data control.