Mac RAM to AI model size: the reference table
- 6 GB runs Nano 4B and 12 GB runs Lite 9B, so 16 GB handles both and loads Quick 26B (24 GB is its comfortable size). 24 GB adds the 27B models, Core and Vision 3.8. 64 GB is where Plus 397B becomes an option (128 GB recommended).
- On disk: Nano is 2.4 GB, Core 27B is 15.1 GB, Vision 3.8 is 15.5 GB. Plus 397B wants 209 GB.
- Speed on my M1 Ultra: Nano hits 71.7 tok/s, Core 27B does 20.7, Plus 397B crawls at 1.59.
- Quick math: usable model size is about your unified RAM minus a few GB for macOS.
Your Mac's unified memory decides which local models you can run. The tables below show what fits in each amount of RAM, each model's size on disk, and the speed I measured on my M1 Ultra. Use them to size up the Mac you own or the one you're about to buy.
RAM → what you can run
| Unified RAM | Typical Mac | Models that run |
|---|---|---|
| 16 GB | MacBook Air | Nano 4B, Lite 9B, Quick 26B (its 16 GB minimum; 24 GB is comfortable) |
| 24 GB | MacBook Pro | + Core 27B, Vision 3.8 27B (tight) |
| 32 GB | MacBook Pro | 27B coding + Vision, comfortably |
| 64 GB | Mac Studio / MBP | All tiers; Plus 397B streams from the SSD at about 1.6 tokens a second (measured, M1 Ultra) and needs 209 GB of free disk |
| 256 GB | Mac Studio | Plus 397B held fully in memory |
Model size and speed (measured, M1 Ultra)
| Model | Params | Disk | Decode tok/s |
|---|---|---|---|
| Nano | 4B | 2.4 GB | 71.7 |
| Lite | 9B | 5 GB | 53.4 |
| Quick | 26B MoE | 15.6 GB | 14.6 (43.6 when re-measured on 2026-08-24, v1.11.804) |
| Core | 27B | 15.1 GB | 20.7 |
| Vision 3.8 | 27B | 15.5 GB | 17.2 (measured 2026-08-19, shipped 1.11.792; see the benchmarks table) |
| Plus | 397B MoE | 209 GB | 1.59 |
Numbers are from an M1 Ultra, MLX 4-bit, batch size 1; the Plus figure is the K=20 sweep of 2026-05-01: 5 prompts of 80 tokens, fresh process. Newer chips (M2/M3/M4) run faster, but the ranking stays the same.
How a 209 GB model runs on a 64 GB Mac
Plus 397B is 209 GB on disk. That's way past 64 GB of RAM, and yet it runs. The trick is that it's a Mixture-of-Experts model, so only a slice of its parameters fire on any given token. Outlier's V9 paged engine holds the active experts in memory and streams the rest off the SSD, which keeps peak memory hovering around 14 GB. See how paged MoE inference works.
Questions about the RAM-to-model-size table
What size AI model can my Mac run?
Roughly your unified RAM minus a few GB for macOS. Nano needs 6 GB and Lite 12 GB, so 16 GB runs both and loads the 26B Quick model (24 GB is its comfortable size), 24 GB adds the 27B models, and 64 GB runs a 397B Mixture-of-Experts model via paged streaming (128 GB recommended).
How fast is local AI on a Mac?
On an M1 Ultra: about 71.7 tok/s for a 4B model, 20.7 for a 27B model, and 1.59 for the 397B model via paged streaming. Newer chips are faster. Speed depends mostly on model size and chip generation.
How much disk do AI models need?
From about 2.4 GB for a 4B model to 209 GB for the 397B model. You only download the tiers you use.
Try Outlier free
Free: Nano + Lite. Pro: $49 once; Black Friday price, all of November: $24.50 once, until 30 November. macOS 15+. Refund window: 30 days, no questions asked.
Download free (276 MB) Buy ProOn your phone?