Apple researchers run large AI models across Mac Studios

Apple researchers have demonstrated running large language models across multiple Mac Studio computers, linking them with MLX and RDMA over Thunderbolt. In their tests, four linked systems ran Muse Spark Open at more than 36 tokens per second, showing how several Macs can share a model that exceeds one machine’s memory.

Watch Desk analysis

What happened

The work evaluates local and distributed inference on M3 Ultra and M5 Ultra Mac Studios, using models including Muse Spark Open, Mistral Medium 3.5 and Qwen 3.8. Apple’s researchers compare pipeline and tensor parallelism, approaches that divide a model’s work across machines. They also report faster time to first token with the M5 Ultra’s Neural Accelerators.

Key findings

• Four Mac Studios, one large model
  Apple reports more than 36 tokens per second for Muse Spark Open across four linked systems.
• More than one way to share the work
  The researchers evaluate pipeline and tensor parallelism for distributed inference.

Why it matters

Running an oversized model locally is not necessarily a matter of finding one enormous computer. Apple’s demonstration suggests a route using multiple Mac Studios, with Thunderbolt links carrying the work between them. That is relevant to developers and researchers exploring local inference, though the reported speed is for this particular setup and model, not a general promise for every Mac or workload.

Our read

This is a practical computing result, not just a diagram of machines holding hands. It gives people considering local AI another configuration to examine, and makes the trade-off between buying one larger system and linking several smaller ones more concrete. Apple’s research is the evidence for the demonstration; the useful next step is seeing how the approach behaves across other models and real workloads.

What to watch

• Whether Apple publishes further performance and scaling results.
• How performance changes with different models, Mac configurations and workloads.
• Whether the software and connection setup are practical for developers beyond this demonstration.

Discussion spark: Would you rather run a large local model across several linked computers, or keep the setup simple with one machine even if that limits the models you can run?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.