Summary: Apple introduced M6 and M5 Ultra on August 25, 2026. M6 is its first 2nm chip, while M5 Ultra is the first quad-die design and goes into Mac Studio. M5 Ultra can carry up to 512GB of unified memory, which Apple says is enough for LLMs with hundreds of billions of parameters on a desk.
Apple used August 25, 2026 to introduce two chips at once: M6 and M5 Ultra. M6 holds the distinction of being Apple’s first chip on a 2nm process, and it arrives in the new Mac mini. M5 Ultra is the company’s first quad-die architecture and its most powerful chip to date, going into the new Mac Studio. Pre-orders for Mac Studio opened that same day, with availability beginning September 22, 2026 in 30 countries and regions. Both chips carry the same message: the next hardware battleground is on-device AI.
M6: 2nm Arrives
M6 combines a 12-core CPU (2 super, 4 performance, 6 efficiency cores, two more than M5) with a 12-core GPU that places a Neural Accelerator in every core. Apple claims the world’s fastest single-threaded performance, along with multithreaded gains of up to 1.2x over M5 and 2.4x over M1. Its Neural Engine is a dual 16-core design that delivers up to 2x the peak AI compute of previous generations, and peak GPU AI compute sits nearly 30% above M5.
The memory setup speaks to mainstream developers: unified memory tops out at 32GB with 170GB/s of bandwidth — 10% more than M5 and 2.5x the M1. If you run local models on a laptop or a small chassis, that bandwidth figure deserves more of your attention than the core count, since token generation for local LLMs is largely a memory-bandwidth problem.
M5 Ultra: Four Dies and 512GB of Memory
For AI work, M5 Ultra is the bigger story. Apple fuses two dual-die M5 Max chips into a single quad-die package using the next generation of UltraFusion, with inter-die bandwidth above 4.4TB/s. The CPU scales to 36 cores (12 super, 24 performance), reaching 1.25x the single-threaded speed and 1.3x the multithreaded speed of M3 Ultra. The GPU scales to 80 cores, again with per-core Neural Accelerators, for up to 4.5x the peak GPU AI compute of M3 Ultra and graphics that run up to 40% faster.
Memory is the number that matters most: up to 512GB of unified memory with 1.2TB/s of bandwidth, 50% more than M3 Ultra. Apple says outright that this is enough to run LLMs with hundreds of billions of parameters on the desk. That is a deliberate shot at the shared-memory ceiling — model weights have to fit in memory to run locally, and 512GB covers model classes that previously implied a multi-GPU server rack. The Mac Studio lineup starts at $2,499 with M5 Max (18-core CPU, up to 40-core GPU, up to 128GB memory, 614GB/s bandwidth) and $5,499 with M5 Ultra, while the 512GB configuration arrives in late October.
A Desk That Becomes a Small Cluster
Clustering is the sleeper feature for AI engineers: over Thunderbolt 5 with RDMA, four Mac Studio units link into a single system and run distributed AI inference up to 3x faster. Add 120Gb/s Thunderbolt 5, Wi-Fi 7 and Bluetooth 6 arriving for the first time through Apple’s N1 chip, support for up to eight displays, and a media engine with hardware AV1 decode and up to 33 simultaneous 8K ProRes 422 streams, and the target user is written on the spec sheet. This is a desktop built around local model workflows — prompt processing for large LLMs, agentic workloads, and local fine-tuning through Core AI, Core ML, Metal, and Xcode.
On-Device AI: The Reality Check
Apple’s pitch stays concrete: faster LLM prompt processing, agentic workloads, and the ability to run — and fine-tune — large models locally, alongside macOS 27 “Golden Gate” and expanded Apple Intelligence features. The strategy presents privacy, a one-time purchase, and no token bills as the alternative to cloud APIs.
The reality check is equally concrete. The 512GB configuration does not ship until late October, and $5,499 is not an entry-level price. The “up to 4.3x AI performance” claim is measured against the previous generation, and turning peak numbers into real local inference quality depends on framework and model optimization that takes time. Quantized open models already run well on this class of hardware; the open question is how much of the cloud-API workflow — tooling, serving, fine-tuning — matures on the desktop too. But the direction is unambiguous: once a flagship desktop can hold a hundred-billion-parameter model in memory, running large models on-device moves from demo to purchase order — and the pressure that places on the cloud inference market over the long run is greater than any keynote slogan.