CryptoGPUs
Special orderShips in 2–5 days
More from Apple6 in AI workstations

Apple Mac Studio M5 Ultra

$4,218.41Was $4,440.44−5%

Black Friday price: you save $222.03 until Dec 1. No code needed.

Paid in Bitcoin, Monero, USDT or another coin you choose.

Free insured shipping. US delivery in about 5 to 11 business days. Shipping details

Overview

The Mac Studio with M5 Ultra has the highest memory bandwidth of any compact local-AI machine here: 1.2TB/s, roughly 4.4 times the 273GB/s of a DGX Spark and almost five times a Strix Halo mini PC. Unified memory starts at 96GB and is offered up to 512GB, all shared by a 30-core CPU and a 64-core GPU in the base model, or up to 36 CPU cores and 80 GPU cores.

Bandwidth is what keeps big models usable. The 96GB base handles 70B models at 4-bit to 6-bit with room for context. The 256GB option fits roughly 235B mixture-of-experts models at 4-bit, and the 512GB option, announced for late October, targets 400B-class models. MacStories measured over 100 tokens per second on an 80B-class mixture-of-experts model on a 256GB unit, and 60 to 85 at 64K to 256K context.

The cost is the main catch. Apple prices the base M5 Ultra at $5,499; with 256GB of memory it is $9,499 on the 30-core chip or $10,799 with the 36-core chip, before storage upgrades. Memory cannot be upgraded after purchase. It runs macOS with MLX, LM Studio and llama.cpp, so software that requires CUDA will not run.

  • M5 Ultra, 96GB or 256GB memory
  • 1.2TB/s memory bandwidth
  • Six Thunderbolt 5 ports, 10GbE
  • Clusters over Thunderbolt 5 RDMA

Specifications

Performance
ChipApple M5 Ultra: 30-core CPU and 64-core GPU base, or 36-core CPU and 80-core GPU
Neural Engine32-core, with a Neural Accelerator in every GPU core
Memory96GB unified (base) or 256GB; 512GB announced for late October
Memory bandwidth1.2TB/s
Storage1TB SSD base, configurable to 2TB, 4TB, 8TB or 16TB
Local AI capacity96GB: 70B at 4-bit to 6-bit. 256GB: about 235B MoE at 4-bit. 512GB: 400B-class MoE
ClusteringThunderbolt 5 with RDMA, up to four Mac Studio systems (Apple claims up to 3x faster inference)
Connectivity
Thunderbolt 54 rear ports plus 2 front ports on M5 Ultra, up to 120Gb/s
USB-A2 rear ports, up to 5Gb/s
VideoHDMI 2.1, up to 8 external displays on M5 Ultra
Ethernet10Gb Ethernet (1, 2.5, 5 and 10Gb)
WirelessWi-Fi 7 (802.11be), Bluetooth 6, Thread via Apple N1
Card slot and audioSDXC UHS-II slot on the front, 3.5mm headphone jack, built-in speaker
Power and size
Dimensions9.5 x 19.7 x 19.7cm (3.7 x 7.7 x 7.7in)
Weight3.6kg (8.0lb) for M5 Ultra
PowerMaximum continuous 480W, 100 to 240V
In the boxMac Studio, power cord
Software
Operating systemmacOS 27 with Apple Intelligence
AI frameworksMLX, Core AI, LM Studio, exo for clustering

Questions and answers

Still unsure? Visit the help center or message us.

How much memory do I need for a 70B model?

The 96GB base model is enough for a 70B model at 4-bit to 6-bit (about 40GB to 58GB) with usable context. Choose 256GB if you want 8-bit or very long context.

Which models need the 512GB configuration?

Mixture-of-experts models around 400B parameters and above at 4-bit, plus long context on them. Apple lists 512GB as an option, and AppleInsider and MacStories report it arrives in late October with no published price yet.

Can I add memory or storage later?

No. Memory is part of the chip package and the SSD is configured at purchase, so choose capacity when ordering.

Does it run CUDA software?

No. It runs Apple's MLX and Metal-based tools such as LM Studio and llama.cpp. Workflows that depend on CUDA, such as some training and video-generation stacks, need an NVIDIA machine.

Can I link several Mac Studios?

Yes. Thunderbolt 5 supports RDMA clustering between systems, and Apple says a four-unit cluster delivers up to 3x faster AI inference than one.

From the blog

All articles
Local AI

Best computer for local AI: DGX Spark vs Mac Studio vs Strix Halo

NVIDIA DGX Spark, Mac Studio M5 Ultra and Ryzen AI Max+ 395 compared on memory, bandwidth, real benchmarks and price, plus when a GPU tower is the better buy.

Pairs well with

Picked to go with this product. Check sizes and connectors for your build before you order.