Mac Studio M5 Ultra Bets on Core AI and MLX for Local Models

The hardware arrived quietly. The bigger change is in the software

Apple chose an odd time to refresh its desktops. It announced the new Mac mini and Mac Studio on August 25 with a press release, held no event, and started shipping both machines on September 2234. The Mac mini comes with either the new M6 chip, which Apple calls its first 2nm part, or the M5 Pro. The Mac Studio moves to the M5 Max and the new M5 Ultra4. Neither enclosure changed, and some outlets called the update a spec bump49.

That label leaves out the most important part of the launch. Apple is pitching the Mac Studio mainly as a machine for running AI models on the device itself, rather than as a video workstation that also does AI. Its launch materials named specific software frameworks, which is unusual for Apple: a brand-new system framework called Core AI, its open-source MLX library, and third-party tools such as LM Studio1114. The silicon got faster, but the more lasting news is that Apple now has a software stack for local AI on the Mac and is selling hardware built around it.

What actually shipped

The base Mac mini with M6 costs $899. It has a 12-core CPU, a 12-core GPU with a Neural Accelerator in every core, and what Apple calls a Dual 16-core Neural Engine. Memory starts at 16GB and goes up to 32GB, with bandwidth of up to 170GB/s1. The M5 Pro Mac mini starts at $1,699, with up to an 18-core CPU, a 20-core GPU, and 64GB of unified memory at 307GB/s1.

The Mac Studio is where the AI pitch is strongest. The M5 Ultra model starts at $5,499 and offers up to a 36-core CPU, an 80-core GPU, and up to 512GB of unified memory at 1.2TB/s of bandwidth. Apple says that combination lets researchers run very large language models entirely on the machine1. This is the first time the Ultra tier has Neural Accelerators in its GPU cores, and Apple cites up to 4.3 times the peak AI compute of the previous generation11. Thunderbolt 5 with Remote Direct Memory Access (RDMA) lets several Mac Studios work together. Apple says a cluster of four delivers up to three times faster inference than a single machine117.

There are a few catches. The 512GB configuration, which is the one most relevant to frontier-scale models, is listed for late October. Macworld notes reports that orders are slipping toward late January2. The M5 Ultra model also costs $200 more than its previous price, which makes it $1,500 more than the previous Ultra cost at its March 2025 launch2. iDrop News adds that the storage is soldered, so buyers have to pick their capacity at purchase8.

Core AI: a new framework at the center

The most notable item for developers is Core AI. Apple describes it as a new framework for building, running, and deploying AI models on Apple silicon. It is designed around unified memory, the CPU, the GPU, and the Neural Engine, so developers can run full-scale LLMs locally and bring their own custom models into apps11. One developer-focused analysis says Core AI automatically decides which part of the chip handles each inference job, whether that's the CPU, the GPU's Neural Accelerators, or the Neural Engine. The same analysis describes it as running alongside Core ML as a successor path, and calls it the layer Apple wants MLX and LM Studio-style apps to use14.

This matters because Apple silicon now has three different places where AI work can run, and until now developers had to decide where each workload should go. A system framework that makes that decision automatically hides the hardware differences. It also gives Apple room to change the hardware underneath in future chips without breaking apps. My view is that Core AI is mainly about keeping developers on Apple's platform over the long term, more than about this year's raw speed.

MLX moves from side project to core tool

Core AI is Apple's official system framework. MLX is the tool the local-AI community already uses. Apple's announcement describes MLX as its open-source machine-learning framework for running, training, and fine-tuning models efficiently on the Mac11. Ars Technica's review says MLX has become well optimized and widely supported by language models and local-AI apps like LM Studio Bionic, in part because the hardware suits these workloads so well13.

MLX's key advantage comes from the chip design. The CPU and GPU share a single pool of memory, so MLX can work on data from either processor without copying it back and forth. That is why one compact Mac can load models that would need several graphics cards on a PC15. On a machine with 512GB of memory, that difference decides which models you can run at all.

The frameworks also connect more closely than the launch materials suggest. According to a June developer guide, Apple opened its Foundation Models framework to custom backends at WWDC 2026. It also shipped MLXLanguageModel, which loads any mlx-community model into the same Swift API, with streaming, tool calling, and structured output20. In practice, a Swift developer can now replace Apple's built-in model with an open-weight model running on MLX through the same API. The M5 Ultra Mac Studio is the hardware that makes that option practical at large model sizes.

Where the coverage agrees, and where it doesn't

Reviewers broadly agree that the Mac Studio is a big step forward for local inference. MacStories called the M5 Ultra a large performance jump for MLX-powered models and titled its review a "dream Mac" for local AI agents12. Its test numbers, summarized elsewhere, show a large mixture-of-experts model generating roughly 84 to 91 tokens per second at short and medium context lengths. The M3 Ultra managed about 45 to 5916. One community tester reported a large overnight jump in batched throughput, which suggests the software was still catching up to the hardware at launch16.

The differences show up in the details, and they decide whether Apple's 4.3x figure means anything in practice. One technical analysis of Apple's own MLX numbers for the base M5 found that the Neural Accelerators made time-to-first-token 3.3 to 4.06 times faster than on the M4. Generation of later tokens improved by only 1.19 to 1.27 times, because memory bandwidth still limits it19. Put plainly, the headline gains mostly show up in how fast the model reads a prompt, not in how fast it writes an answer. Users running agents with long contexts will feel the difference. People who mostly chat with a model will notice less.

The comparison with Nvidia is just as mixed. MacStories found that an RTX 5090 PC beat the M5 Ultra at every model size the PC could finish. The M5 Ultra was still closer to the 5090 than to the M3 Ultra12. One tester found that a pair of Nvidia DGX Spark units outperformed a single M5 Ultra on one model16. Gadget Review reminded readers that Apple's performance claims had not been independently verified at launch6. These results point the same way. Apple wins on memory capacity: one machine with 512GB can hold models that don't fit on any single consumer GPU. Nvidia still wins on speed for models small enough to fit in its memory14.

The most lasting limitation is software, not hardware. MLX, llama.cpp's Metal backend, and Core AI handle inference and LoRA fine-tuning. They do not cover the large body of research code and custom kernels written for Nvidia hardware19. Core AI can make Apple's own platform easier to use, but it does nothing to bring CUDA-based code to the Mac.

Why it matters

The Mac mini is a solid mainstream update. Apple pitches the M6 as capable of "agentic AI workflows" at an $899 starting price, alongside the Apple Intelligence and Siri AI features in macOS 271. The Mac Studio is the clearer signal of where Apple is heading. Apple now has a layered stack. Core AI sits at the system level, MLX serves as the open framework for research and fine-tuning, and Foundation Models gives app developers an API that can now run open-weight models1120. The new hardware is the first designed specifically to run that stack.

Overall, Apple is not trying to beat Nvidia on speed. It is building a different option: quiet, private, high-memory machines with an AI software stack it controls from the chip up to the Swift API. Two things will show whether that works. One is whether the 512GB Studios actually ship on time. The other is whether developers adopt Core AI or keep calling MLX and llama.cpp directly. For now, the frameworks are a more important part of this launch than the new enclosures-that-weren't.