Apple shows a giant open-weight model running across four Mac Studios.

On October 15, Apple's machine-learning team published work on distributed inference with MLX. In a featured test, it ran Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters, across four M5 Ultra Mac Studios. Apple reports 36.73 output tokens per second using NVFP4 quantization at an 8,192-token context.

The performance figure is measured under a particular setup and is Apple's own reporting. It does not describe how quickly a full interactive application would respond under multiple users, nor does it compare the total cost with rented infrastructure.

Very large open-weight models have typically been associated with expensive server installations. Demonstrating that such systems can be split across developer-owned machines broadens the conversation about who can operate powerful models, how data can be handled and where inference happens.

At the same time, local ownership creates costs that cloud pricing can conceal in different ways: initial purchases, depreciation, power, orchestration, upgrades and downtime. The right answer may vary dramatically by workload and data sensitivity.

Cortist's central question is not whether local AI can be impressive—it clearly can—but when it is the better operational choice. A company with a stable high-volume workload and strict data controls may value ownership differently from one with irregular demand.

The evaluation should hold model quality constant and compare total cost per accepted task, rather than look only at token throughput. A test also needs to account for the impact of quantization and realistic concurrency.