By memory size
Local LLMs for a Mac
with 32GB of memory
On a Mac with 32GB of memory, models up to 16GB run comfortably alongside other apps: 21 of the 33 models measured here. The smartest is Qwen3.5-9B (84% on knowledge, 98% on math, 6.5GB). "Measured" here also counts the models measured for the series page, beyond the top page's text lineup.
Other sizes: 8GB · 16GB · 24GB · 32GB · 36GB · 48GB or more
To find your Mac's memory, choose the Apple menu → About This Mac (for example, 16 GB).
Picks by use
| Use | Model | Rule |
|---|---|---|
| If unsure, start here | gpt-oss:20b | The smartest that replies within 5 seconds (1 more within the margin) |
| Smartest (if you can wait) | Qwen3.5-9B | The highest share of right answers (7 more within the margin) |
| Replies right away | GLM-4-9B-0414 | The smartest that replies within 1 second |
| English → Japanese translation | gemma-4-12B-it | The highest translation score (9 more within the margin) |
Runs comfortably (up to 16GB)
| Model | Memory | Reply time | Knowledge | Math | EN→JA |
|---|---|---|---|---|---|
| Qwen3.5-9B | 6.5GB | about 40.3 s | 84% | 98% | — |
| gemma-4-12B-it | 8.1GB | about 7.5 s | 85% | 97% | 35 pts |
| gpt-oss:20b | 12.8GB | about 1.8 s | 83% | 98% | 33 pts |
| Ornith-1.0-9B | 6.2GB | about 3.8 s | 85% | 94% | — |
| Qwen3-14B | 10.6GB | about 8 s | 82% | 96% | 34 pts |
| Qwen3.5-4B | 3.5GB | about 29.7 s | 81% | 96% | — |
| LFM2.5-24B-A2B | 14.9GB | about 7.1 s | 81% | 96% | 28 pts |
| NVIDIA-Nemotron-Nano-12B-v2 | 8.1GB | about 6.8 s | 80% | 94% | 31 pts |
| Ornith-1.5-9B | 6.4GB | about 2.5 s | 77% | 94% | 33 pts |
| gemma-4-E4B-it | 5.4GB | about 2.6 s | 79% | 92% | 32 pts |
| NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 | 6.5GB | about 2.6 s | 76% | 93% | 34 pts |
| Qwen3-4B | 3.9GB | about 3.1 s | 72% | 93% | 29 pts |
| MiniCPM5-2B | 2GB | about 1.5 s | 73% | 90% | 23 pts |
| DeepSeek-R1-0528-Qwen3-8B | 6.4GB | about 9.4 s | 72% | 89% | 23 pts |
| LFM2.5-2.6B | 2GB | about 2.3 s | 73% | 79% | 29 pts |
| NVIDIA-Nemotron-3-Nano-4B-BF16 | 3.3GB | about 1.9 s | 72% | 80% | 24 pts |
| LFM2.5-8B-A1B | 5.4GB | about 1.7 s | 71% | 78% | 29 pts |
| GLM-4-9B-0414 | 6.7GB | about 0.1 s | 72% | 77% | 32 pts |
| LFM2-24B-A2B | 14.8GB | about 0.1 s | 53% | 66% | 32 pts |
| LFM2.5-1.2B-JP-202606 | 1GB | about 0.2 s | 51% | 63% | 33 pts |
| DeepSeek-R1-Distill-Qwen-1.5B | 1.5GB | about 1.7 s | 23% | 21% | 3 pts |
Runs with heavy apps closed
More than half the memory, but up to 20.8GB should run. Close heavy apps other than a browser first.
| Model | Memory | Reply time | Knowledge | Math | EN→JA |
|---|---|---|---|---|---|
| qwen3.8:27b | 18.2GB | about 7.7 s | 87% | 100% | 31 pts |
| gemma4:26b | 18.7GB | about 3.2 s | 85% | 98% | 34 pts |
| muse-glimmer:30b | 17.6GB | about 15 s | 83% | 98% | 36 pts |
| glm-4.7-flash:latest | 19.4GB | about 7.5 s | 76% | 95% | 32 pts |
| gemma3:27b | 17.5GB | about 0.4 s | 76% | 91% | 33 pts |
| qwen3-coder:30b | 19.4GB | about 0.1 s | 69% | 89% | 31 pts |
| Hy-Memory-30B-A3B | 19.2GB | about 0.1 s | 66% | 88% | 31 pts |
Image and video models (guide)
Models whose measured peak memory is up to 20.8GB. Close other apps first to be safe.
- Z-Image Turbo (Image, 18.5GB)
- FLUX.2 [klein] 4B (Image, 16.7GB)
Apple list price, tax included (guide)
The tax-included list price is unconfirmed. The buy-page terms forbid automated fetching, so this site did not retrieve yen amounts.
| Memory | Apple list price, tax included (guide) | Difference from the size below |
|---|---|---|
| 8GB | Unconfirmed | — |
| 16GB | Unconfirmed | Unconfirmed |
| 24GB | Unconfirmed | Unconfirmed |
| 32GB | Unconfirmed | Unconfirmed |
| 36GB | Unconfirmed | Unconfirmed |
| 48GB or more | Unconfirmed | Unconfirmed |
Source: Apple’s official buy pages (www.apple.com/jp/shop/buy-mac/mac-mini / www.apple.com/jp/shop/buy-mac/mac-studio). Checked 2026-10-24. No outbound link.
About this guide
- Memory was measured on a Mac Studio (M5 Max, 36GB). Memory use is about the same on any Apple silicon Mac, so it carries over.
- Reply times are from this Mac. Slower chips (a MacBook Air, say) reply more slowly.
- "Comfortable" means the model takes at most half the memory; "with heavy apps closed" up to 65%. The picks follow the same rules as the text page, among the models that fit.
- What "9B", "A3B" or "it" in the names mean: how to read a model name.
Task recommendations and capabilities · Measurements by model