Apple's New M6 and M5 Ultra Chips Could Revolutionize Local AI
M5 Ultra Mac Studio offers 4.3x faster AI performance than M3 Ultra.
1Apple's New Chip Generation
Apple unveiled new M6 and M5 Ultra chips for Mac mini and Mac Studio. The M6 is Apple's first chip manufactured using 2nm process technology. Its dual 16-core Neural Engine delivers 2x faster AI performance than previous generations. The 12-core GPU features Neural Accelerators in each core for the first time. According to Apple, these innovations provide 4x faster AI performance compared to M4-based Mac mini.
2M6 Mac Mini's AI Potential
The base M6 Mac mini (16GB RAM) is expected to generate over 60 tokens/sec with models like Qwen 3.5-35B-A3B, significantly surpassing the ~17 tokens/sec of M4-based Mac mini. However, the M6 Mac mini lacks Thunderbolt 5 support, which is exclusive to the M5 Pro model. This limitation prevents clustering multiple Mac minis for AI workloads.
3Power of M5 Ultra Mac Studio
The M5 Ultra Mac Studio delivers 4.3x higher AI performance than M3 Ultra and 9.8x more than M1 Ultra. With 512GB unified memory and 1.2TB/s memory bandwidth, it enables running massive AI models entirely on-device with privacy. The 512GB RAM M5 Ultra Mac Studio models will be available in late October.
4Thunderbolt 5 and Clustering Capability
M5 Ultra supports Thunderbolt 5, allowing multiple Mac Studios to cluster with shared memory. Apple claims a cluster of four Mac Studios delivers 3x the performance of a single unit. This feature provides significant advantages for large-scale AI model inference tasks.
5Leap in Token Generation Speeds
While M3 Ultra Mac Studio averages 35 tokens/sec with DeepSeek-V4-Flash, M5 Ultra is expected to reach 120 tokens/sec. This speed surpasses many cloud-based AI chat services' 'fast modes'. Only specialized NVIDIA PC clusters or inference providers like Cerebras can outperform this performance.
6Running Large Models Locally
M5 Ultra enables running traditionally quantization-heavy large models. For example, the 744B parameter GLM-5.2 model runs at 17 tokens/sec on M3 Ultra. M5 Ultra is expected to handle such massive models at even higher speeds, making frontier open-weight models feasible on local hardware.
7Role of MLX Framework
Apple's MLX machine learning framework maximizes the AI potential of M-series chips. It enables local AI agents to continuously scan and automatically categorize project data. These capabilities have revolutionary potential for content creators and researchers' workflows.
8Future of Local AI
Apple's advancements with M-series chips hint at the future of local AI. While requiring over $20,000 investment, achieving data-center-level AI performance with a single Mac Studio is now possible. These developments reinforce Apple's leadership in AI hardware while paving the way for privacy-focused local AI solutions.
Original article: MacStories — Apple'ın Yeni M6 ve M5 Ultra Çipleri Yerel AI'da Devrim Yaratabilir