Apple introduced two desktops this week, sending M6 into the new Mac mini and M5 Ultra into the new Mac Studio. The Apple M6 is the company’s first 2 nm desktop part and targets the mainstream desktop, while M5 Ultra is a quad-die design with up to 512GB of unified memory for high-capacity local AI and professional compute. Apple published the full announcement alongside the launch. We are a bit behind, just given our massive Hot Chips 2026 coverage.
Apple M6 and M5 Ultra #
The Apple M6 brings a 12-core CPU split into two super cores, four performance cores, and six efficiency cores, joined by a 12-core GPU with a Neural Accelerator in each core and a dual 16-core Neural Engine. Unified memory tops out at 32GB with up to 170GB/s of bandwidth. Apple claims 1.2x faster multithreaded CPU performance than M5, nearly 30 percent more peak GPU AI compute, and 10 percent more memory bandwidth. Apple also says the Dual 16-core Neural Engine delivers up to 2x peak compute over prior generations and that both engines can run at once for faster model execution. We called the Apple Mac Mini M4 the standard for compact desktops in earlier coverage, and this refresh carries that mini form factor forward.
The M5 Ultra is the larger of the pair. Apple UltraFusion ties two dual-die M5 Max chips into one quad-die SoC, with an inter-die fabric Apple rates above 4.4TB/s. Scaling reaches a 36-core CPU with 12 super cores and 24 performance cores, an 80-core GPU with Neural Accelerators, and a 32-core Neural Engine. Unified memory still supports up to 512GB but is served by 1.2TB/s of bandwidth, which Apple says is 50 percent higher than the M3 Ultra. Apple claims up to 1.3x faster multithreaded performance than the M3 Ultra, again from preproduction parts. Another fun note is that the M5 Ultra has a PCIe Gen6 SSD, which is faster than the M.2 SSDs currently on the market.
We called the prior M3 Ultra Mac Studio the big local AI play. That capacity holds large language models that common discrete GPUs cannot fit, and the faster 1.2TB/s pool should improve tokens-per-second for on-device inference. Higher bandwidth and an updated GPU hopefully mean that this is notably better. We are focused on the M5 Ultra, but there is an M5 Max version as well with up to an 18-core CPU, 40-core GPU, and 128GB of memory.
Final Words #
Pricing has certainly increased. An Apple Mac Studio M3 Ultra 512GB we purchased for around $10K, but the 256GB M5 Ultra model is over $11K. Apple says the new M5 Ultra 512GB will be available in October, so for now you can only order the 256GB model. It will be interesting to see how much that one costs.
Something worth considering is that, from a desk-side AI perspective, the NVIDIA GB300 Station line is about $100K but offers an NVIDIA GPU with several times higher memory bandwidth and more total system memory. We had a Qwen3.6-35B-A3B NVFP4 Fast model running at well over 26K tokens/s at c=512, which is wild to see.
Local AI is getting really cool.