Local AI Gets Serious on the M5 Ultra Mac Studio
Summary
Federico Vitici reviews the new M5 Pro Mac mini and M5 Ultra Mac Studio, focusing on their local AI performance. He benchmarks models like Kimi K3 and Qwen 3.8, highlighting massive improvements in token pre-fill and generation speeds over the M3 Ultra. The episode also compares the Mac Studio to an NVIDIA 5090 PC, noting the Mac's advantages in unified memory, size, and quiet operation despite the GPU's raw speed.
Topics discussed
Intro: Mac Mini and Mac Studio reviews
Hardware specs and review unit details
Focus on local AI performance testing
Mac Mini M5 Pro vs M4 Pro benchmarks
John's M4 Pro Mac Mini setup workflow
Sponsor: Astroad Workbench remote desktop
Mac Studio M5 Ultra specs and GPU architecture
Running local agents for 99 days on M3 Ultra
M5 Ultra memory bandwidth and prefill speeds
Benchmark visualizations and interactive charts
Running Qwen 3.8 Flash Next locally
Quantization levels and RAM usage analysis
Future potential: 512GB RAM and Thunderbolt 5
Daily driver apps and voice model integration
Sponsor: Claude from Anthropic
Comparison with NVIDIA RTX 5090 GPU
Mac Studio vs PC: Size, noise, and preference
Outro and App Stories Plus teaser
Listen ad-free on Castria