Mac-1 — 6.6B Local AI Agent for macOS

Source: Threads post by @aditya.rao310 / X/Twitter by @cjzafir

What It Is

Mac-1 is a 6.6B parameter model (fine-tuned on Qwen 3.5 series) that runs entirely offline on any Mac. Built by @cjzafir with a full native macOS application layer.

Key Specs

SpecValue
Model size6.6B parameters (4B + 2B dense)
RAM required7GB (12GB ideal)
Speed~65 tok/s with reasoning ON
macOS tools487 native tools
QuantizationQ4_K_M — 99% accuracy
Training data670M handcrafted tokens, 113K+ rows
ConnectivityWorks fully offline

What It Can Do

Chains together native macOS tools to accomplish multi-step tasks from a single instruction:

  • Schedule a meeting
  • Set a reminder
  • Draft a follow-up email
  • All in one instruction, no internet needed

Why It Matters

  • Truly native macOS app, not a web wrapper
  • Offline-first — no cloud dependency
  • Tool chaining — 487 tools composed into multi-step workflows
  • Prefill + KV cache optimized for low-latency inference

This is the direction of on-device AI agents — small, fast, deeply integrated with the OS.