Mac mini Guide
Choosing and sizing Mac minis for OpenClaw: one machine or several, gateway vs inference vs workers.
Answer a few questions on the home page to get setup suggestions tailored to your budget and goals.
One Mac mini or several?
Multiple Mac minis are often needed. Running the OpenClaw gateway, a local model (Ollama/MLX), and several agents on the same machine is not recommended unless you have a high-spec box (e.g. M4 Pro with 48–64 GB). On a single mid-range machine, you’ll hit memory and CPU contention: the gateway, the model, and agent tasks all compete for RAM and bandwidth. For anything beyond light, single-session use, plan to split roles across machines.
- Gateway-only (one Mac): channels + cloud API only. Any M1/M2/M4 with 8GB is fine.
- Gateway + local model on one machine: only sensible for a single agent and a modest model (e.g. M4 24GB + 14B), or if you have an M4 Pro 48GB+ “beast.”
- Recommended for local models + multiple agents: one Mac as brain/gateway, another as dedicated inference node (Ollama/MLX), and optional worker Mac minis for task execution. See roles below.
Roles: gateway (brain), inference node, workers
Splitting these across Mac minis keeps the gateway responsive and lets you size each machine for its job.
Gateway / brain
Runs the OpenClaw Gateway (Node.js), channels, session manager, and agent orchestration. Light on compute; can optionally run a small local fallback. Sweet spot: M4 24 GB (~$999) or M4 Pro 48 GB if this box also runs a local model.
Inference node
Dedicated to Ollama or MLX — serves the LLM so the gateway and workers stay snappy. Memory bandwidth is the bottleneck. Best value: M4 Pro 48–64 GB (273 GB/s bandwidth); base M4 32 GB works for smaller models.
Swarm workers
Execute tasks (shell, browse, tools) delegated by the brain. Don’t need large RAM for inference. M4 16 GB (~$599) per worker is ideal; at ~4W idle, several workers are cheap to run 24/7.
Gateway-only vs gateway + local model (on one machine)
Gateway only
OpenClaw gateway + channels (WhatsApp, Telegram, etc.). No local LLM. Any M1/M2/M4 Mac mini is fine; 8GB RAM is enough.
Gateway + local model on same machine
Only for light use or a high-spec machine (M4 Pro 48GB+). RAM is the limit: 16GB → 7–8B, 24GB → 14B, 32GB → 27B–32B. For multiple agents or heavier inference, use a separate inference node.
RAM vs what you can run (Ollama, Apple Silicon)
| Unified memory | Example models |
|---|---|
| 8GB | Llama 3.1:8b Q4, Gemma 2B, Phi 3 mini. Good for light use. |
| 16GB | Llama 3.1:8b Q8, Qwen2.5:14b Q4, Gemma 9B. Sweet spot for quality/speed. |
| 24GB | Gemma 27B Q6, Llama 3 70B Q2. Better reasoning, slower. |
| 32GB+ | Larger 27B–70B at higher quants. M2 Pro Mac mini or M4 16GB+. |
When to add another Mac mini
Start with one machine; add more as you hit limits. Entry: single M4 24 GB runs gateway + 14B local fallback for light use. Mid-range: M4 Pro 48 GB as brain/inference + M4 16 GB as worker (~$2,500). Cluster: 1× M4 Pro 64 GB + multiple M4 16 GB workers for parallel agents and larger models. See Architecture for BRAIN + nodes and networking.
M1 vs M2 vs M4
All run OpenClaw and Ollama. M2 and M4 have faster memory bandwidth and better GPU; for 14B+ models, M2 16GB or M4 16GB is a noticeable step up from M1 16GB. For gateway-only, M1 is enough. For inference-heavy setups, M4 Pro’s 273 GB/s bandwidth gives roughly 2× the token throughput of the base M4.
Storage
Models can be 4GB–40GB+ each. 256GB internal is tight if you run several models; 512GB or external SSD gives room. See Hardware & Storage for DRAM vs DRAMless and external drive picks.