// Mac Device Check

What can your Mac
actually run?

Most people know their Mac by name — MacBook Air, MacBook Pro — not by chip or memory. This page bridges the gap. Pick your device and memory; get a straight answer.

> The constraint on Mac is RAM, not chip generation. Here's what yours unlocks.
iPhone / iPad Mac Linux

Not sure which model? Check Apple menu → About This Mac

Find it at Apple menu → About This Mac — look for "Memory" or "Unified Memory"

info On Mac, memory (RAM) is the key variable — two MacBook Pros from the same year can land in completely different tiers. Chip generation matters less than you'd think; unified memory architecture means even older M1 chips use their memory efficiently for AI.

// Why Mac is different from iPhone

The Mac AI story nobody tells you

iPhone AI gets all the coverage. Mac can go significantly further — and for different reasons than you might expect.

// No jetsam
Apps don't get killed silently
iOS enforces a per-app memory budget and kills apps silently when models exceed it. macOS doesn't. If memory gets tight, it slows down via SSD swap — the model keeps running.
→ Slower, not dead
// Unified memory
CPU, GPU, and Neural Engine share one pool
Apple Silicon's unified memory means the GPU can work on model weights that were just used by the CPU — no copy overhead. This is genuinely better for LLM inference than discrete VRAM setups.
→ Memory works harder here
// SSD as fallback
Models larger than RAM can still load
When a model exceeds physical RAM, macOS pages layers to the SSD instead of refusing to load. A 7B model on 8GB RAM is slow (~5 tok/s) but functional. On iPhone, it crashes.
→ 8 GB Macs have more reach than you'd expect
// Background inference
Generate while you work
On macOS, a model can generate in the background while you write in another window. On iPhone, switching apps suspends inference. Mac lets the model finish its thought.
→ Real multitasking
// Context depth
Much longer conversations
Without a per-app memory cap, you can push 32K–128K token context windows with the right models. That means feeding an entire codebase or document into a single prompt and getting coherent answers.
→ Whole-project reasoning
// Network serve
Serve your iPhone and iPad too
With a local inference server (Ollama, or Mulberry IDE's network mode), your Mac can serve AI responses to your iPhone and iPad over Wi-Fi — one model, multiple devices, still private.
→ Your Mac as a local AI hub

// Find your RAM

Not sure how much memory you have?

Most people can't name their RAM without looking it up — that's normal. Apple puts it exactly one click away.

  • 01 Click the Apple menu () in the top-left corner of your screen
  • 02 Select About This Mac
  • 03 Look for "Memory" or "Unified Memory" — it will say something like 16 GB
  • 04 Come back and select that number in Step 2 above

// The four Mac tiers

What your memory unlocks

Entry
8 GB Unified Memory
MacBook Air M1/M2/M3 · Mac mini M1/M2 · iMac M1/M3 (base)
The most misunderstood tier. 4B-class models (Phi-4 mini, Qwen3 4B, Gemma 3 4B) run at full speed with headroom to spare. 7B models load and run too — via SSD swap at ~5 tok/s. Useful for patient tasks; not the right choice for interactive writing or coding.
Phi-4 mini · 2.2 GB ★ Qwen3 4B · 2.5 GB ★ Gemma 3 4B · 2.5 GB Llama 3.2 3B · 2 GB Llama 3.1 8B · 4.7 GB ⚠ slow via swap
Standard
16–18 GB Unified Memory
MacBook Air M1/M2/M3/M4 · MacBook Pro M1/M2 14" · Mac mini M2/M4 · iMac M3/M4
The everyday workhorse. Llama 3.1 8B is your default — full Neural Engine speed (~15–25 tok/s on M2+), long sessions without slowdown. Gemma 3 12B adds a quality step for complex reasoning. 18 GB configs (M3 Pro) handle Qwen 2.5 14B cleanly.
Llama 3.1 8B · 4.7 GB ★ Gemma 3 12B · 7 GB ★ Qwen 2.5 7B · 4.5 GB Mistral 7B · 4.5 GB Qwen 2.5 14B · 8 GB ⚠ tight on 16 GB
Pro
24–36 GB Unified Memory
MacBook Air M2/M3/M4 (max) · MacBook Pro M3 Pro / M4 Pro · Mac mini M4 Pro
Serious AI hardware. Qwen 2.5 32B and Gemma 3 27B run cleanly — this is where local AI starts competing with cloud API quality. 36 GB configs (M3 Pro, M4 Pro) can also push Llama 3.1 70B at Q2 quantization.
Qwen 2.5 32B · ~20 GB ★ Gemma 3 27B · ~15 GB ★ Qwen 2.5 14B · 8 GB Llama 3.1 70B Q2 · ~30 GB ⚠ 36 GB only
Max
48 GB+ Unified Memory
MacBook Pro M3 Max / M4 Max · Mac Studio M2/M4 Max / Ultra · Mac Pro
No practical ceiling. Llama 3.1 70B Q4 runs at speed. Multiple large models simultaneously. Serve inference to other devices on your network. At 192 GB+ (Mac Studio Ultra, Mac Pro), 405B-class models become possible.
Llama 3.1 70B Q4 · ~40 GB ★ Qwen 72B · ~42 GB ★ Mistral Large · ~40 GB Multiple simultaneous models 405B Q4 · ~200 GB ⚠ 192 GB+ only
lock

Whatever tier your Mac runs, the model lives on your machine. No cloud round-trip, no account, no telemetry. Your prompts, your documents, and your code never leave the device.