Overview
Atlas is an M1 Ultra Mac Studio (128GB unified memory) running as a private inference appliance — a dedicated local AI server that handles model inference, agent orchestration, and automation pipelines across a personal fleet spanning two properties in New York.
The core philosophy: own your compute, own your data. No API keys for everyday inference. No cloud dependency for things that can run locally.
Hardware
- Machine: Apple Mac Studio, M1 Ultra
- Memory: 128GB unified memory
- Storage: 500GB ClawWorkspace (APFS), Clawdrive-01 for nightly backup
- Network: Tailscale tailnet, accessible from anywhere
Model Stack
Atlas runs a tiered model stack tuned for different workload types:
- Primary: Gemma 4 26B MoE — fast, efficient, everyday inference
- Fallback: Qwen 3.5 35B — subagent worker tasks
- Heavy synthesis: Qwen 122B — deep research and long-context work
- Planned: DeepSeek V4-Flash (284B/13B active) for the heavy slot
All models served via a custom gateway with Telegram bot interface.
What It Runs
- Hermes — primary agent, LaunchDaemon managed, Telegram pipeline
- Ollama — model serving and management
- Fleet HQ — live dashboard on port 28400, React/Express, mobile-responsive
- MTAPI — NYC subway service on port 8000
Fleet Role
Atlas sits at the center of a broader personal fleet:
- Rosie (M1 Ultra) — worker/secondary inference
- Sterling (M4 Pro Mac Mini) — personal agent machine
- RedFoxy + UpperWesty (M1 Minis) — property nodes
- PrintClaw (M1 Mini) — Bambu X1C 3D printer controller
All machines connected via Tailscale. Nightly rsync, git autobackup at 3AM.
Why It Matters
Running local inference at this scale changes how you interact with AI. Latency drops to milliseconds. Privacy is absolute. The model stack is yours to tune. Atlas is the reason every automation in the fleet can call a capable model without a network request leaving the building.