What is the Apple M6 chip?

M6 is Apple's first 2nm chip, announced August 25, 2026. It has a 12-core CPU, a 12-core GPU with Neural Accelerators, a Dual 16-core Neural Engine, and up to 170GB/s of unified memory bandwidth — two more CPU and GPU cores than M5. It debuts in the Mac mini starting at $899 with 16GB of memory and 256GB of storage, which Apple says delivers up to 4x faster AI performance than the prior generation.

What is the M5 Ultra and how much memory does it support?

M5 Ultra is Apple's first quad-die chip, announced August 25, 2026 for the Mac Studio. It scales to a 36-core CPU, an 80-core GPU, up to 512GB of unified memory and 1.2TB/s of memory bandwidth. The base Mac Studio M5 Ultra is $5,499 with a 30-core CPU, 64-core GPU, 96GB of memory and 1TB of storage. Apple claims up to 4.3x faster AI performance than the previous Mac Studio.

When do the new Mac mini and Mac Studio ship?

Both go on sale September 22, 2026. Configurations with 512GB of unified memory are not available until late October 2026. The Mac Studio update comes 16 months after the previous refresh in April 2025.

Is a Mac Studio M5 Ultra good for running local LLMs?

For model capacity, it is exceptional — 512GB of unified memory at 1.2TB/s fits models that no consumer NVIDIA card can hold. The limitation is ecosystem, not silicon. Ollama and LM Studio run well on Apple Silicon, but CUDA-dependent agent stacks do not. Perplexity's Portable Computer, launched the same day, is NVIDIA-only on Linux with no macOS support announced.

Quick Answer

What Is Apple M6 and M5 Ultra? AI Specs and Prices

Published: August 26, 2026

The Short Answer

On August 25, 2026, Apple announced two chips: M6, its first 2nm processor, and M5 Ultra, its first quad-die design. They debut in a refreshed Mac mini (from $899) and Mac Studio (M5 Max from $2,499, M5 Ultra from $5,499), both on sale September 22, 2026.

For local AI the headline number is not the process node. It is 512GB of unified memory at 1.2TB/s on a maxed M5 Ultra.

M6: The 2nm Chip

Spec	M6
Process	2nm (Apple’s first)
CPU	12-core
GPU	12-core, with Neural Accelerators
Neural Engine	Dual 16-core
Memory bandwidth	Up to 170 GB/s
vs M5	+2 CPU cores, +2 GPU cores
AI performance	Up to 4× faster (Apple’s claim)

M6 lands in the Mac mini at $899 with 16GB unified memory and 256GB storage — a $200 increase over the prior generation, with base memory held at 16GB. The M5 Pro Mac mini starts at $1,699 (24GB / 512GB), scaling to an 18-core CPU, 20-core GPU and up to 64GB.

The “Dual 16-core Neural Engine” is the notable structural change — two NPU blocks rather than one, aimed squarely at on-device model execution.

M5 Ultra: The Quad-Die Part

Spec	M5 Ultra
Architecture	Quad-die (Apple’s first)
CPU	30-core base, up to 36-core
GPU	64-core base, up to 80-core
Unified memory	96GB base, up to 512GB
Bandwidth	1.2 TB/s
Base price	$5,499 (96GB / 1TB)
AI performance	Up to 4.3× faster than prior Mac Studio

The Mac Studio also offers M5 Max from $2,499 (36GB / 512GB). Apple additionally claims 2× faster storage and up to 1.3× faster CPU. This refresh arrives 16 months after the April 2025 update.

Availability caveat that matters: the 512GB configurations — the entire reason a local-AI buyer would look at this machine — do not ship until late October 2026, a month after everything else.

Why 512GB Of Unified Memory Is The Real Story

Unified memory means the GPU addresses the full pool. On a discrete-GPU PC, a model must fit in VRAM: NVIDIA’s RTX PRO 6000 Blackwell tops out at 96GB, and DGX Spark at 128GB. A maxed M5 Ultra offers 512GB — four times the Spark, over five times the RTX card.

That capacity gap decides what runs at all. Very large open-weight models that simply cannot be loaded on a single NVIDIA workstation card fit comfortably in a Mac Studio.

Two honest counterweights:

Bandwidth still favours NVIDIA at the top. RTX PRO 6000 runs 1.8TB/s against Apple’s 1.2TB/s. Token generation is memory-bound, so the RTX card generates faster — on models small enough to fit.
Capacity you can load is not capacity you can afford. Apple’s memory upgrades are famously expensive, and 512GB is nowhere near the $5,499 entry price.

The Ecosystem Limit — Read This Before Buying

Apple’s silicon is not the constraint. CUDA is.

The clearest illustration arrived the same day. On August 25, 2026, Perplexity launched Portable Computer, a fully local agent harness with an OS-enforced sandbox and zero per-token cost for local steps. It runs on NVIDIA DGX Spark, RTX PRO and GeForce RTX with 24GB+, on Linux. There is no macOS version, and none announced.

So on the same day Apple shipped the best local-AI memory configuration available, the most significant local-AI agent product shipped without Mac support.

What does work well on Apple Silicon: Ollama and LM Studio, both mature on Metal/MLX, both able to exploit unified memory. If your workflow is “run open-weight models locally,” a Mac is excellent. If it is “run this specific CUDA-dependent agent framework,” a Mac cannot do it at any memory size.

Who Should Buy What

Mac mini M6 ($899) — general use with occasional local inference. 16GB base is tight for serious model work; budget for an upgrade. The 4× AI claim is real but relative to a low starting point.

Mac mini M5 Pro ($1,699) — up to 64GB makes mid-sized local models genuinely practical. The value pick of this launch.

Mac Studio M5 Max ($2,499) — the sensible local-LLM machine. Enough memory for most open-weight work without $5,499+ commitment.

Mac Studio M5 Ultra ($5,499+) — only if you need to hold very large models in memory and you are committed to the Metal/MLX ecosystem. Wait for the late-October 512GB configurations if capacity is the reason you are buying.

None of the above — if you are buying to save on API costs. At $0.041 (Gemini 3.7 Flash) to $0.275 (Claude Opus 5) per 30K-in/5K-out task, a $5,499 Mac Studio represents an enormous number of frontier calls that are more capable than anything you will run locally.