Mac Studio Large-Language-Model Tests Put M3 Ultra Through Its Paces, Look Ahead to M5 Ultra

Apple’s Mac Studio M3 Ultra uses a unified memory architecture, allowing its CPU and GPU to share up to 256GB of memory and reducing data-transfer bottlenecks when large language models are deployed locally. For developers who need to run models such as Llama 3.1 without continuing to pay for cloud computing, the ability of a single machine to handle large models is a key measure of value.

The latest tests examine the Mac Studio M3 Ultra’s inference performance on large language models including Llama 3.1. They also assess a technical approach using clusters of Macs to exceed the 256GB per-machine memory limit and handle larger models. The report also looks ahead to an M5 Ultra that Apple may release in 2026, although no firm launch date, pricing or official specifications are available.