AppStories
AppStories

Local AI Gets Serious on the M5 Ultra Mac Studio

September 21, 2026

AI Summary

5 min read

Local AI Gets Serious on the M5 Ultra Mac Studio

Federico Viticci spent 99 days this summer running local AI agents around the clock on an M3 Ultra Mac Studio to support his iOS 27 review. Every five minutes, an agent would wake up to process new research, categorize it, cross-reference it with his notion database, and reconcile it with existing chapters. It worked, but it was slow — a single categorization could take ten minutes. The machine could handle the workload, but the experience of waiting for responses made real-time interaction impractical. Now, with the M5 Ultra Mac Studio in hand, Viticci reports that the local AI experience has fundamentally changed.

The Hardware That Makes It Possible

The review unit Viticci tested is an M5 Ultra with an 80-core GPU, 256GB of unified memory, and 4TB of storage. Apple's key architectural change in the M5 generation is placing a neural accelerator inside each GPU core — so the Ultra has 80 of them. Combined with a memory bandwidth jump from 819 GB/s on the M3 Ultra to 1.2 TB/s (a 50% increase), the machine can move data in and out of its shared memory pool much faster.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

What you'll learn

  • 1 (00:11) **Introduction and Hardware Overview** - John and Federico introduce the episode, which covers the M5 Pro Mac Mini and the M5 Ultra Mac Studio, with a focus on local AI performance.
  • 2 (02:50) **Mac Studio Specs and the Local AI Use Case** - Federico outlines the M5 Ultra Mac Studio's specs (80-core GPU, 256GB RAM) and his prior year of using a Mac Studio as a local AI server.
  • 3 (05:12) **Memory Bandwidth and Performance Jumps** - Federico details the M5 Ultra's 50% higher memory bandwidth (1.2 TB/s vs. 819 GB/s) and its impact on token prefill and generation speeds.
  • 4 (08:40) **The Flash Next Model and Real-World Agent Speed** - Federico describes running Qwen 3.8 Flash Next (a hybrid architecture model with a 125B parameter model and a 51B parameter lookup table) on the M5 Ultra.
  • 5 (10:10) **Quantization and Memory Management** - Federico discusses running different quantizations of Flash Next on the M5 Ultra, including 4-bit, 5-bit, 6-bit, and 8-bit versions.
  • 6 (11:55) **The 512GB Model and Future Possibilities** - Federico speculates on the upcoming 512GB M5 Ultra and its potential for running even larger models.
  • 7 (12:54) **Practical Agentic Setup and Voice Integration** - Federico explains that he now exclusively uses local models on the M5 Ultra with apps like Hermes Agent and Open Minis, and the speed makes interactive use viable.

+ Full timestamped outline available in the app

Show Notes

This week on AppStories, John interviews Federico about his review of the M5 Ultra Mac Studio and it local AI capabilities.

On AppStories+, Federico takes listeners behind the scenes on how he tested the M5 Ultra Mac Studio.

Also available on YouTube here.

Sponsored By:

  • Workbench: Remote desktop for AI agents and headless Mac minis. Free to use for 20 minutes/day.
  • Claude – For problems worth solving — get started with Claude today.

Links and Show Notes

Running Local AI Models on Apple's Latest Macs

Apple Hardware

Local Models

Local Inference and Clustering

Agent Apps and Tools

Frontier Models and Workflow