Apple refreshed the Mac mini and Mac Studio today with two new chips built explicitly around local AI inference: the M6, its first 2nm system-on-a-chip, and the M5 Ultra, which tops out at 512GB of unified memory and 1.2TB per second of bandwidth. The M6 Mac mini starts at $899 with 16GB of memory, and the M5 Ultra Mac Studio starts at $5,499. Preorders open today, with shipping on September 22, 2026, though the 512GB Studio configuration won't ship until late October.
There are no major new features on either machine — the industrial design carries over and the story is entirely in the silicon. Apple is positioning the refresh around a use case that didn't exist when these desktops were first engineered: developers and researchers daisy-chaining Mac minis and Mac Studios to run open-weight language models locally, as an alternative to renting time on Nvidia GPUs.
The M6 has a 12-core CPU with two of what Apple calls super cores, four performance cores, and six efficiency cores — the first Apple SoC to use all three core types. Apple claims up to 40% faster multi-threaded CPU performance versus the M4 from two generations ago. The GPU picks up two additional cores for a 12-core total, and unified memory bandwidth reaches 160GB per second, though memory tops out at 32GB.
Key facts
- 01The M6 Mac mini starts at $899 with 16GB of memory; the M5 Ultra Mac Studio starts at $5,499 and scales to 512GB of unified memory.
- 02M5 Ultra hits 1.2TB per second of unified memory bandwidth with 36 CPU cores and 80 GPU cores — effectively two M5 Maxes on one die.
- 03M6 is Apple's first 2nm chip, with a 12-core CPU Apple claims is up to 40% faster in multi-threaded workloads than the M4.
- 04Preorders open today, shipping September 22, 2026; the 512GB Studio config slips to late October 2026.
- 05The refresh follows macOS 26.2, shipped December 2025, which enabled distributed MLX inference over Thunderbolt 5.
That 32GB ceiling is why serious local-inference workloads land on the M5 Ultra. The Ultra is effectively two M5 Max dies fused into one SoC, yielding 36 CPU cores (12 super, 24 performance) and 80 GPU cores. Its 1.2TB per second of unified memory bandwidth and 512GB memory ceiling are the specs that matter for anyone loading a large open-weight model into RAM in one shot.
“low-latency communication between Thunderbolt 5 hosts for use cases including distributed AI inference using MLX”— Apple, macOS 26.2 release notes
The distributed-inference use case became mainstream on Mac after macOS 26.2 shipped in December 2025. The release enabled distributed MLX inference over Thunderbolt 5, letting multiple Macs pool their unified memory across a wired link. MLX is Apple's open-source array framework tuned for the M-series' unified memory architecture, and it has become the connective tissue for hobbyists and professional developers running models that no single consumer device could hold.
The economics behind the shift are straightforward. Frontier coding agents like Claude Code and Codex run on cloud models that charge by the token, and heavy users are already questioning whether the monthly bill is sustainable. Open-weight releases from Qwen and DeepSeek can handle a large slice of the same coding and reasoning workloads at zero marginal token cost — the trade is that the user pays in local compute and electricity instead.
The gap that remained was hardware. A standard MacBook Pro can run smaller open-weight models, but the coding-agent-quality tier of Qwen and DeepSeek variants generally needs far more memory than a laptop carries. Chaining several Mac minis or Studios over Thunderbolt 5 has become the workaround, and Apple's new specs — particularly the M5 Ultra's 512GB ceiling — are pitched directly at that community.
Beyond the AI story, both machines pick up standard upgrades. Storage speeds hit up to 15GB per second, roughly double the previous generation. The Mac mini ships with 2.5Gb Ethernet as standard with an optional 10Gb upgrade. Both use Apple's N1 chip for Wi-Fi 7 and Bluetooth 6. Mac mini configurations with M5 Pro start at $1,699, and Mac Studio with M5 Max starts at $2,499.
Apple also confirmed the desktops will ship with macOS 27 Golden Gate, suggesting the annual OS update is close to release for the rest of the Mac lineup. There are no third-party benchmarks yet, so Apple's 40% multi-threaded claim on the M6 remains a vendor number until independent testing lands. The M5 Ultra's practical inference throughput against a similarly priced Nvidia workstation is also an open question — memory bandwidth is one axis, and raw matrix throughput is another.
The strategic read is that Apple is quietly building an on-ramp to compete with Nvidia in a market segment Nvidia currently owns by default. The company isn't shipping a datacenter GPU or courting hyperscalers. It's optimizing a mass-market desktop line for the specific workflow of a developer who wants to run a 200B-parameter open model without a cloud bill, and pricing the entry point at $899. If open-weight models keep closing the gap with frontier cloud APIs, the value of a $5,499 Mac Studio with 512GB of unified memory grows with every release — and that's a market position Nvidia's consumer cards don't cleanly answer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
