Apple announces Mac Mini and Mac Studio models with M6 and M5 Ultra chips for local AI development and inference

Summary

Apple has refreshed the Mac mini and Mac Studio with the M6 (the first 2 nm M-series chip) and M5 Ultra (dual M5 Max, 512 GB unified memory). The company is explicitly positioning this hardware for local AI inference and development workloads. Thanks to Thunderbolt 5 daisy-chaining enabled by macOS 26.2, developers can connect multiple Macs to run open-weight models like Qwen and DeepSeek without relying on the cloud.

Why it matters

This move is a direct investment in the growing market of developers who want to run and fine-tune LLMs locally. Apple’s unified memory architecture offers a cost advantage per token compared to renting GPU cloud instances. The Thunderbolt 5 distributed inference pathway unlocked by macOS 26.2 creates a modular local AI stack that could fundamentally change how mid-sized teams handle inference tasks without a single GPU server costing over $30,000.

Highlights

• M5 Ultra: 36 CPU + 80 GPU cores, 1.2 TB/s memory bandwidth, up to 512 GB unified memory
• M6, despite being Apple’s first 2 nm M-series chip, is limited to 32 GB of unified memory
• Thunderbolt 5 + MLX enables daisy-chaining multiple Macs for distributed local inference

Critical take

The M6’s 32 GB memory cap effectively disqualifies it for serious local LLM workloads out of the box. The 'distributed inference' promise still requires multiple premium Macs to compete with a single high-end GPU node, suggesting the cost advantage over the cloud is narrower than marketing claims imply. The article reads like a press release summary, labeling the M5 Ultra a 'true power beast' without providing independent benchmark data.