Mac mini with M6
Best compact Apple choice for a quiet always-on agent gateway plus modest local inference; choose 32GB if local models matter more than basic gateway work.
Compatibility and purchase evidence →Local AI buying guide
Choose hardware for the workload, not the sticker.
Compatibility, memory, thermals, model fit, and official purchase sources for local runtimes and agent gateways.
Last checked: 2026-08-26 · Manufacturer links only · no fabricated benchmarks
Start with the architecture
Ollama and LM Studio run models; Open WebUI provides a UI; Hermes Agent and OpenClaw orchestrate tools and channels. A gateway can run on a small low-power host while inference runs on a separate machine or cloud provider. Do not buy a GPU just to host a lightweight gateway.
Local-engine watch (checked 2026-09-02): Perplexity’s official blog describes Lily as a specialized Apple silicon engine for Qwen3.6-35B-A3B used in Perplexity Computer hybrid compute. It is an open-source demo, not a directory SKU. Official demo notes Apple GPU family 10 or later (M5 and newer). See the Change Radar and official Lily post.
Mac mini with M5 Pro
The most compelling new Mac mini tier for local models plus always-on agent services, if the 48GB/64GB configurations fit the budget.
Compatibility and purchase evidence →Mac Studio with M5 Max or M5 Ultra
Choose only when large local models or heavy concurrent workloads justify the cost; otherwise a 32GB/64GB Mac mini or discrete-GPU PC is easier to justify.
Compatibility and purchase evidence →NVIDIA RTX 5090 desktop workstation
Best choice for CUDA-heavy local AI and generation workflows when power, heat, and desktop size are acceptable.
Compatibility and purchase evidence →Raspberry Pi 5 (8GB or 16GB) gateway
Buy for an inexpensive always-on gateway, not for serious local model inference.
Compatibility and purchase evidence →Cross-stack decision matrix
| Hardware | Best use | Memory guidance | Evidence |
|---|---|---|---|
| Mac mini with M6 compact desktop | Best compact Apple choice for a quiet always-on agent gateway plus modest local inference; choose 32GB if local models matter more than basic gateway work. | 16GB standard; configure 24GB or 32GB for local models and concurrent services. | View evidence → |
| Mac mini with M5 Pro compact pro desktop | The most compelling new Mac mini tier for local models plus always-on agent services, if the 48GB/64GB configurations fit the budget. | Up to 64GB unified memory according to Apple announcement; prefer 48GB or 64GB for larger local models and multiple services. | View evidence → |
| Mac Studio with M5 Max or M5 Ultra high-memory pro desktop | Choose only when large local models or heavy concurrent workloads justify the cost; otherwise a 32GB/64GB Mac mini or discrete-GPU PC is easier to justify. | Apple announcement: up to 128GB M5 Max and 512GB M5 Ultra. Configure for the model class and concurrency actually planned. | View evidence → |
| NVIDIA RTX 5090 desktop workstation discrete-GPU desktop | Best choice for CUDA-heavy local AI and generation workflows when power, heat, and desktop size are acceptable. | 32GB dedicated VRAM; system RAM of 64GB is a sensible workstation target for local AI, containers, datasets, and browser/agent overhead. | View evidence → |
| Raspberry Pi 5 (8GB or 16GB) gateway low-power ARM gateway | Buy for an inexpensive always-on gateway, not for serious local model inference. | 4GB is enough for a simple gateway; 8GB is a comfortable target; 16GB is unnecessary for most gateway-only deployments. | View evidence → |
How to use this guide
- Identify whether you need a gateway, local inference, or both.
- Choose the exact model family and quantization before selecting RAM or VRAM.
- Leave headroom for the operating system, context/KV cache, containers, browser automation, and concurrent services.
- Check official runtime support for the OS, chip, driver, and version.
- Buy only after testing a representative task or reserving a return path.