Frontier-class AI, running on your desk.
Open-weight models matched to your Apple Silicon — owned outright, fully offline, and kept current as better models ship. No API keys, no usage meters, no one else reading the traffic.
The right model for the machine you own.
| Hardware | RAM | Model range | Capability | Typical installs |
|---|---|---|---|---|
| Mac mini M4 Pro | 24 GB | Up to ~32B | Fast daily assistant, coding copilot, summarization | Qwen 3 32B, Phi-4 14B, Gemma 3 27B |
| Mac Studio M4 Max | 48–128 GB | Up to ~70–80B | Frontier-class chat, reasoning, document analysis | Llama 3.3 70B, Qwen 2.5 72B, DeepSeek R1 70B |
| Mac Studio cluster | 192 GB+ | 100B+ / multi-model | Run several models simultaneously or the largest MoE | Llama 4 Scout 109B, parallel agents |
Models we install and keep current.
| Model | Best at | Size | Min hardware |
|---|---|---|---|
| Llama 3.3 70B | All-round workhorse — chat, reasoning, writing | ~70B | 48 GB+ |
| Qwen 2.5 72B | Structured tasks, multilingual, instruction following | ~72B | 48 GB+ |
| DeepSeek R1 70B | Math, logic, step-by-step reasoning | ~70B | 48 GB+ |
| Llama 4 Scout 109B | Frontier reasoning (MoE — fast for its size) | ~109B MoE | 64 GB+ |
| Qwen 3 32B | Best quality at 24 GB — daily driver | ~32B | 24 GB+ |
| Qwen 2.5 Coder 32B | Code generation, refactoring, debugging | ~32B | 24 GB+ |
| DeepSeek R1 32B | Reasoning specialist, lighter hardware | ~32B | 24 GB+ |
| Gemma 3 27B | Balanced chat + vision, strong on-device | ~27B | 24 GB+ |
| Phi-4 14B | Snappy assistant — great speed-to-quality ratio | ~14B | 16 GB+ |
| Gemma 3 4B | Ultra-light, instant responses, 8 GB machines | ~4B | 8 GB+ |
Sizes are approximate (quantized weights). Strengths are qualitative — we test on your hardware before install. This lineup changes as better open models ship.
Curated for your work, kept current.
Matched to your hardware
We benchmark every model on the actual Mac you own — not synthetic scores. You get the largest, highest-quality model that runs comfortably on your silicon.
Swapped as better models ship
Open-weight AI moves fast. When a new release outperforms what you're running, we update your stack — same workflow, better results.
Always offline
Every model runs entirely on your hardware. No API calls, no telemetry, no cloud fallback. Your prompts and your data never leave the box.
Not sure which model fits your hardware?
We assess your machine, match the right stack, and install it — ready the same day.
Book a consultation