Which local model fits your Mac?
The answer is a number, not an opinion. A model has to sit in your unified memory alongside macOS, your other applications and the context it is generating into. Here is the arithmetic, what it means for every Mac Apple currently sells, and the one mistake that costs people the most.
The rule
Keep the model file under 60 percent of your installed memory. Treat 80 percent as a wall you do not touch.
The file size on disk is very close to what the weights occupy in memory, so you can read it straight off the download. What the remaining 40 percent pays for is the KV cache that grows as the context fills, whatever else you have open, and the memory macOS itself needs. Past 80 percent the machine starts swapping to disk, and a model that swaps does not run slowly, it stops being usable.
| Unified memory | Comfortable file size | Absolute ceiling | Realistic model |
|---|---|---|---|
| 8 GB | up to 4.8 GB | 6.4 GB | a 7B to 9B at 4-bit |
| 16 GB | up to 9.6 GB | 12.8 GB | a 9B at 6-bit, or a 14B at 4-bit |
| 18 GB | up to 10.8 GB | 14.4 GB | a 14B at 5-bit |
| 24 GB | up to 14.4 GB | 19.2 GB | a 14B comfortably, a 30B MoE at low quant |
| 32 GB | up to 19.2 GB | 25.6 GB | a 30B class model at 4-bit |
| 48 GB | up to 28.8 GB | 38.4 GB | a 30B at 6-bit, or a 70B at low quant |
| 64 GB | up to 38.4 GB | 51.2 GB | a 70B at 4-bit |
| 128 GB | up to 76.8 GB | 102.4 GB | a 70B at 8-bit, or a 120B at 4-bit |
These are the same thresholds Slate applies when it labels a model comfortable, tight or too big before you download it. The app reads your actual installed memory rather than asking you to look it up.
Why not 100 percent
Unified memory is shared between the CPU and the GPU, which is what makes Apple silicon good at this in the first place: there is no copy across a bus. But the GPU cannot claim all of it. On a 24 GB Mac the practical Metal working set we measure is around 19 GB, and the 80 percent line above lands at 19.2 GB. That is not a coincidence, it is the same ceiling approached from two directions.
Which is why a model that reports 21 GB on a 24 GB machine will load, appear to work, and then produce its first token a minute later. Nothing is broken. The weights are being paged off an SSD.
The mistake that costs the most
Choosing parameter count over quantisation.
A 30B model squeezed into 2-bit will fit where a 9B at 6-bit fits, and it will usually be worse. Aggressive quantisation damages the weights faster than the extra parameters recover, and the damage shows up exactly where you notice it: instruction following, arithmetic, and staying coherent over a long answer.
A serviceable heuristic: 4-bit is the floor for real work, 5-bit and 6-bit are where quality stops being the bottleneck, and 8-bit is rarely worth double the memory. Pick the largest model that fits comfortably at 4-bit or better — not the largest model that technically loads.
| Label | Bits | What to expect |
|---|---|---|
| Q8_0 | 8 | Near-lossless, and roughly double Q4 for gains you will struggle to notice. |
| Q6_K | 6 | The quality plateau. If it fits comfortably, take it. |
| Q5_K_M | 5 | Excellent balance. The usual right answer on 16 to 24 GB. |
| Q4_K_M | 4 | The floor for serious work. Noticeably weaker at long reasoning. |
| IQ4_XS | ~4 | A smarter 4-bit scheme. Often how a large MoE is made to fit at all. |
| Q3, Q2 | 3, 2 | Emergency measures. Prefer a smaller model at a higher bit depth. |
What your Mac has
Apple silicon ships in fixed memory configurations, so the answer for your machine is one row. Hold Option and choose About This Mac, or read it off the table.
| Chip | Memory options | What runs well |
|---|---|---|
| M1, M2, M3 | 8, 16, 24 GB | A 14B at 16 or 24 GB; at 8 GB it is 7B territory |
| M4 | 16, 24, 32 GB | 14B comfortably, 30B MoE at 32 GB |
| M1, M2, M3, M4 Pro | 16 to 48 GB | 14B to 30B depending on the configuration |
| M1 to M4 Max | 32 to 128 GB | 30B comfortably, 70B from 64 GB |
| M1 to M3 Ultra | 64 to 512 GB | 70B at high precision and beyond |
Slate needs macOS 26 on Apple silicon. There is no Intel build, and on an Intel Mac none of the above applies because there is no unified memory to share.
Images and voice need their own room
An image model is a second set of weights. Running one alongside a chat model on 16 or 24 GB means both are competing for the same pool, which is why Slate offloads parts of the image pipeline to the CPU on smaller machines rather than pretending they fit. Speech models are small enough not to matter — a few hundred megabytes, not gigabytes.
If you plan to use both at once, budget for the larger of the two and treat the 60 percent line as applying to the pair.
Or let the app do the arithmetic.
Slate reads your Mac's memory, labels every model comfortable, tight or too big before you download it, and on first run suggests the largest one that fits.