Verdict
ZeroClaw on a Mac Mini M4 Pro 48 GB is the local-LLM endgame at small form factor in 2026. Llama 3.3 70B Q4 at 9-10 tok/s, ZeroClaw's offline RAG over a documents folder, network egress denied at iptables — this is what 'self-hosted AI without the cloud' looks like at its peak in mid-2026.
Setup notes
Asahi Linux if you want the Linux experience (with caveats — server-grade not yet production). Or macOS-host with Ollama + ZeroClaw on top. Configure ZeroClaw network egress allowlist to nothing. Mount your documents folder as read-only into the agent's RAG store.
Performance
Idle: 12W. Sustained 70B Q4 inference: 35-45W. RAG indexing: one-time, scales with corpus size.
What breaks
- Cloud-LLM workflows by design (this is the point)
- Anything requiring frontier-tier reasoning
Want to know more
See the full ZeroClaw review and the Mac Mini M4 / M4 Pro buyer's notes.