Apple unveils M6 and M5 Ultra chips with 2nm silicon and massive local AI power

Apple has unveiled two new bits of silicon at opposite ends of the desktop: M6 for the new Mac mini and M5 Ultra for the new Mac Studio. One is aimed at everyday users and developers; the other is designed to chew through giant video projects, 3D renders and local AI models without immediately asking for a cloud subscription.

The M6 is Apple’s first chip built using a 2-nanometre process, while M5 Ultra is the company’s first M-series chip to use a quad-die architecture. In less brochure-friendly language, Apple is shrinking the distance between a compact desktop and a workstation, then adding a frankly silly amount of memory to the expensive one.

M6 brings 2nm to the Mac mini

M6 features a 12-core CPU with two super cores, four performance cores and six efficiency cores. Apple says it delivers the world’s fastest single-threaded performance, up to 1.2 times the multithreaded performance of M5, and up to 2.4 times the performance of M1.

The chip also includes a 12-core GPU, with a Neural Accelerator in every core, plus a Dual 16-core Neural Engine. Apple says the new Neural Engine offers up to twice the peak compute of previous generations, while the GPU provides nearly 30 per cent more peak AI compute than M5.

Unified memory bandwidth rises to up to 170GB/s, with support for up to 32GB of unified memory. That matters for on-device AI because the CPU, GPU and Neural Engine can share the same memory pool instead of constantly shuffling data between separate hardware islands.

M5 Ultra is the monster under the desk

M5 Ultra is built by connecting two dual-die M5 Max chips through Apple’s UltraFusion technology. The result is a quad-die system on a chip with an up-to-36-core CPU, an up-to-80-core GPU and a 32-core Neural Engine.

Apple claims up to 1.2TB/s of unified memory bandwidth, 50 per cent more than M3 Ultra, and up to 512GB of unified memory. That is enough to keep very large datasets and AI models in local memory rather than repeatedly sending work to a remote server.

The GPU includes a Neural Accelerator in each core, alongside second-generation Dynamic Caching, hardware-accelerated mesh shading and third-generation ray tracing. Apple says graphics performance is up to 40 per cent faster than M3 Ultra, while peak GPU compute for AI is up to 4.5 times higher.

Built for local AI and serious creative work

Apple is positioning M6 at students, developers, AI hobbyists and everyday creative users, with examples including Photoshop, Xcode and the game Mixtape. M5 Ultra is aimed at 3D rendering, visual effects, scientific analysis, high-resolution video editing and running large AI models locally.

The new Mac Studio’s media engine includes dedicated H.264, HEVC and ProRes encode and decode hardware, as well as AV1 decoding. Apple says the M5 Ultra can handle complex timelines with multiple streams of high-resolution video and effects.

Apple Intelligence features are currently available for testing through its beta programme and are expected to arrive with macOS 27 this autumn on supported devices and languages. Apple notes that availability may vary by region and language.

Apple’s numbers need a footnote

The headline performance figures come from Apple’s own August 2026 testing using preproduction Mac mini and Mac Studio systems. The comparisons use selected industry-standard benchmarks, so they are useful indicators rather than independent measurements of every workload.

That distinction matters particularly for AI. More memory and bandwidth can make running a large model locally much easier, but the actual experience still depends on the model, software optimisation and whether the application can use Apple’s frameworks effectively.

Our opinion

M6 is the more interesting announcement for normal humans. A 2nm chip, a bigger CPU and GPU, and a second Neural Engine could make the Mac mini a seriously capable little box for coding, creative work and local AI without turning the desk into a server room.

M5 Ultra, meanwhile, is Apple doing what Apple does best: taking an already powerful workstation chip, joining more of it together and giving it enough memory to make cloud AI look less inevitable. The numbers are enormous, but so will be the price of the Mac Studio built around them.

The real win is not a vague promise of “AI compute”. It is the ability to keep more demanding work on the machine in front of you, where your data is faster to access and does not need to take a scenic route through someone else’s data centre. Apple still needs to publish full machine pricing and independent tests, but these chips look like a meaningful step rather than a routine badge swap.