Apple unveils refreshed Mac Studio with M5 Max and a new M5 Ultra aimed at on-device AI

Apple on Tuesday introduced a new Mac Studio workstation built around its next-generation M5 Max and an all-new M5 Ultra chip, positioning the machine as a step-change for on-device artificial intelligence and high-end creative and data workloads. The company said the new system targets professionals who want to run large AI models locally, without relying on cloud processing.

Apple’s announcement says the Mac Studio with M5 Max delivers up to 4.3x faster AI performance versus the prior generation, alongside faster storage and graphics and a higher ceiling for compute-heavy tasks. The M5 Max configuration includes an 18-core central processing unit (CPU) and a GPU up to 40 cores, with “Neural Accelerators” built into each GPU core, and supports up to 128GB of unified memory, a shared high-bandwidth pool of memory used across the CPU, GPU, and AI components.

At the top end, the Mac Studio with M5 Ultra scales to a 36-core CPU, up to an 80-core GPU, and up to 512GB of unified memory. Apple also described the M5 Ultra as enabling “enormous LLMs” (large language models) to run entirely on-device. The company tied its AI performance improvements to Neural Accelerators that speed up matrix multiplication, a core operation in neural networks.

In addition to raw compute, Apple emphasized memory bandwidth and connectivity. For M5 Ultra configurations, Apple cited up to 1.2TB/s of unified memory bandwidth, and for the overall Mac Studio line it highlighted a new connectivity set: Wi‑Fi 7 and Bluetooth 6 arrive on Mac Studio for the first time, along with Thunderbolt 5 for high-speed external storage, expansion, and hubs. Apple said Thunderbolt 5 also supports clustering multiple Mac Studio systems, and it described the result as faster distributed AI inference across systems.

Apple’s AI message extends beyond hardware to the software and frameworks it says developers and researchers rely on. It introduced “Core AI,” described as a new framework for building, running, and deploying AI models on Apple silicon, designed around unified memory and the CPU, GPU, and Neural Engine. Apple also highlighted MLX, its open-source machine learning framework optimized for Apple silicon, and said the combination of those tools with Xcode supports an end-to-end workflow from experimentation to deployment.

For scaling out local inference, Apple said teams can cluster multiple Mac Studio units using built-in Thunderbolt 5 support for RDMA (remote direct memory access), a technique that allows systems to access each other’s memory with reduced overhead. Apple said a cluster of four Mac Studio systems can deliver up to 3x faster AI inference than a single system, and described the clustering approach as creating a shared memory pool that can load larger open-weight models.

Apple also detailed other professional performance features it is including in the Mac Studio hardware. It cited third-generation hardware-accelerated ray tracing for more realistic lighting and faster 3D and VFX workflows, enhanced shader cores for smoother real-time viewport navigation, and a Media Engine that supports hardware-accelerated video codecs including H.264, HEVC, ProRes, and AV1. The company also noted that the M5 Ultra’s Media Engine has twice the video encode and decode blocks as M5 Max and said it can play up to 33 streams of 8K ProRes 422 at 30 fps on M5 Ultra.

The company tied the product launch to its software platform and personal assistant features, pointing to macOS 27, which Apple says includes “Siri AI” and broader Apple Intelligence capabilities across apps. Apple described Siri AI as drawing on personal context to find information across messages, emails, and photos, answering questions, and taking actions across applications, with systemwide integration in areas like Spotlight and context menus. It also said users can ask Siri about what is on their display using “Visual Intelligence.”

Apple did not provide pricing in Tuesday’s announcement, nor did it disclose the specific Mac Studio configurations and shipping timelines beyond availability. The company also did not quantify how much of the AI performance delta depends on software optimization versus chip-level changes, though it offered multiple comparative figures for AI compute, CPU speed, storage, and graphics across prior-generation chips. For buyers focused on enterprise workflows, the announcement similarly does not specify supported cluster configurations beyond the four-system example or outline any management tooling for multi-node inference.

Looking ahead, the key watch items are how developers adopt Apple’s new “Core AI” framework and MLX for local model deployment, and whether professional software and model runtimes align quickly with the new memory ceilings and Thunderbolt 5 clustering approach. Apple said pre-orders begin immediately and availability starts September 22, setting a near-term window for hands-on testing of local AI inference performance and real-world throughput on complex creative and data workloads.