Last updated: 26 August 2026.
On 25 August Apple announced the new Mac Studio with M5 Max and M5 Ultra, claiming up to 4.3x the peak AI compute of the M3 Ultra. If you run local language models, or are budgeting for a machine that will, the obvious question follows: does the Neural Engine, the AI chip that has been sitting in every Apple processor for years, finally join in?
The short answer: no. Ollama, MLX, llama.cpp and LM Studio compute on the GPU, on the M5 Ultra just as before. The Neural Engine stays idle while a local language model answers, and that is unlikely to change. The longer answer explains where Apple's 4.3x actually comes from and what to look at before buying. Most coverage of the announcement skips this distinction entirely.
What Apple announced
The M5 Max offers an 18-core CPU, up to 40 GPU cores, up to 128GB of unified memory and 614GB/s of memory bandwidth. The M5 Ultra joins two Max dies: up to 36 CPU cores, 80 GPU cores, up to 512GB of memory and 1.2TB/s of bandwidth, roughly half again more than the M3 Ultra. Orders opened on 25 August, deliveries start 22 September; the 512GB configuration follows in October. US pricing starts at $2,499 for the M5 Max and $5,499 for the M5 Ultra.
For local AI, one footnote is the real story: the M5 Ultra is the first Ultra chip with Neural Accelerators in every GPU core. To see why that matters more than any TOPS figure, you have to separate two things Apple's marketing sells under one label.
Two different pieces of silicon
The Neural Engine is a discrete block next to the CPU and GPU, 16 cores, in iPhones since 2017 and in Macs since the M1. It is built for fixed, predictable compute patterns, the kind that vision, speech and embedding models have, and it is extremely power efficient at them. The only way to reach it is Apple's Core ML layer.
The Neural Accelerators are something else: matrix compute units embedded in every single GPU core since the M5 generation, comparable to the tensor cores on an Nvidia card. They are addressable directly through Metal, which happens to be exactly the path every LLM runtime already uses. Apple's MLX framework has been using them since late 2025.
When Apple says 4.3x AI performance, it means the GPU with its new accelerators. The Neural Engine contributes nothing to that number.
Why the runtimes skip the Neural Engine
Three reasons, and none of them is developer laziness.
First, access: the only road to the Neural Engine runs through Core ML. A model has to be converted, loses flexible input lengths on the way, and must live with the quantization formats Core ML supports. Ollama and llama.cpp are built on an open model format and Metal; that foundation cannot be casually ported to Core ML.
Second, the workload: generating an answer token by token is not bound by compute but by memory bandwidth. For every single token, all model weights travel through memory once. The Neural Engine and the GPU share the same unified memory bandwidth, so the Neural Engine cannot, even in principle, be faster here. Its advantage is efficiency, not speed, and efficiency matters more in a phone on battery than in a Mac Studio on mains power.
Third, shapes: the Neural Engine wants fixed tensor shapes. A chat with a growing context is the exact opposite.
That it can be done regardless is documented by developer Alvaro Videla, who wrote the whole path up in a freely readable book, The Apple Neural Engine Inference Book: conversion from GGUF, INT8 as the safe baseline, splitting the model into chunks of roughly 250MB at most, and KV cache management through Core ML's state machinery. It is the most thorough public documentation of this route, and precisely for that reason the most honest answer to why nobody takes it in daily practice: every one of those steps is manual work, per model.
The chip is not exactly wasted, by the way. Dictation, photo analysis and Apple's own on-device model of about 3 billion parameters run right there, as does speech recognition like WhisperKit. The Neural Engine gets used every day, just not for chatting with a large language model. Apple's answer to that gap was not to open up the Neural Engine, but to move matrix units to where LLM software already computes: into the GPU.
Where the new performance actually lands
The Neural Accelerators mainly speed up prompt processing, the ingestion of context before the first token of the answer appears. Apple quotes up to 4x faster LLM processing versus the M3 Ultra, and for the M5 Max up to 10.7x faster prompt processing versus the M1 Max. These are vendor figures with the usual caveats, the same ones we walked through in the Rapid-MLX case. But the direction matches the architecture: prompt processing is compute bound, which is exactly where matrix units help.
In practice: if you feed long documents, run RAG over company files, or hand a coding assistant large amounts of source as context, you will see the biggest gains. The speed of the answer itself, tokens per second, improves roughly in line with memory bandwidth, so about half again over the M3 Ultra, not fourfold.
What this means for a buying decision
For running local language models the order of criteria is: memory capacity first, because the model has to fit in full. Memory bandwidth second, because it sets the answering speed. GPU compute third, for prompt processing. The TOPS figure of an NPU, whether Apple's, AMD's or Qualcomm's, says almost nothing about LLM suitability as long as no runtime uses it. Selecting hardware by that number means buying past your workload.
A running M3 Ultra does not get worse with this announcement; software still decides the speed there. If you are buying new and plan on large models or long contexts, the M5 Ultra is the first machine that combines 1.2TB/s and the accelerators in a box that scales to 512GB.
Whether such a machine pays off in your case at all, measured against confidentiality, request volume and the alternatives, is the question to settle before the order, not after. We work through it with your own tasks in a pilot project, measured rather than guessed. An overview of what local AI does for mid-sized companies is here.
Frequently asked questions
Do Ollama, LM Studio or MLX use the Apple Neural Engine?
No. Every mainstream LLM runtime on the Mac computes on the GPU through Metal. The Neural Engine is only reachable through Apple's Core ML layer, requires fixed tensor shapes and a separate model conversion, and it could not speed up token generation anyway, because that step is bound by memory bandwidth, not compute. While a local language model is answering, the Neural Engine sits idle.
What is the difference between the Neural Engine and the Neural Accelerators?
The Neural Engine is a separate 16-core block that has shipped in iPhones since 2017 and in Macs since the M1. The Neural Accelerators are new with the M5 generation and live inside every single GPU core. They are reachable directly through Metal, and therefore through MLX, with no Core ML detour. Apple's headline figure of 4.3x AI performance comes from these GPU units, not from the Neural Engine.
What does the Mac Studio M5 Ultra change for local LLMs in practice?
Three things, in this order: up to 512GB of unified memory, so large models fit at all. 1.2TB/s of memory bandwidth, roughly half again more than the M3 Ultra, which translates directly into tokens per second. And Neural Accelerators in an Ultra chip for the first time, which mainly speed up prompt processing, meaning the ingestion of long documents and context.
Can you run an LLM on the Neural Engine at all?
Yes, but only with considerable effort through Core ML: conversion from GGUF, INT8 quantization, splitting the model into chunks of roughly 250MB at most, and managing the KV cache by hand. Alvaro Videla documents the complete path in his ANE book. Apple itself runs its own on-device model of about 3 billion parameters this way. For chat-style use of larger models, the effort does not pay off.