Can I run an LLM on my Mac?
Apple Silicon shares memory between the CPU and GPU. By default macOS lets the GPU use about 67% of RAM on machines under 64 GB, and about 75% on 64 GB and larger — less than the 95% we assume for a discrete GPU with the same advertised memory. The tables below use those default limits.
Macs we track
| Mac | Unified memory | Usable (default) | Models that fit (8K) |
|---|---|---|---|
| Mac M1 16GB | 16 GB | 10.7 GB | 165 |
| Mac M4 16GB | 16 GB | 10.7 GB | 165 |
| Mac M2 24GB | 24 GB | 16.1 GB | 199 |
| Mac M3 24GB | 24 GB | 16.1 GB | 199 |
| Mac M4 Pro 24GB | 24 GB | 16.1 GB | 199 |
| Mac M1 Pro 32GB | 32 GB | 21.4 GB | 202 |
| Mac M4 32GB | 32 GB | 21.4 GB | 202 |
| Mac M3 Pro 36GB | 36 GB | 24.1 GB | 202 |
| Mac M4 Pro 48GB | 48 GB | 32.2 GB | 207 |
| Mac M1 Max 64GB | 64 GB | 48.0 GB | 215 |
| Mac M4 Max 64GB | 64 GB | 48.0 GB | 215 |
| Mac M4 Pro 64GB | 64 GB | 48.0 GB | 215 |
| Mac M2 Max 96GB | 96 GB | 72.0 GB | 219 |
| Mac M3 Max 128GB | 128 GB | 96.0 GB | 227 |
| Mac M4 Max 128GB | 128 GB | 96.0 GB | 227 |
| Mac M2 Ultra 192GB | 192 GB | 144.0 GB | 245 |
| Mac M3 Ultra 256GB | 256 GB | 192.0 GB | 248 |
| Mac M3 Ultra 512GB | 512 GB | 384.0 GB | 278 |
How many models fit at each memory size
| Unified memory | Usable (default) | Models that fit | Largest that fits | Example machines |
|---|---|---|---|---|
| 16 GB | 10.7 GB | 165 | Trinity Mini | Mac M1 16GB, Mac M4 16GB |
| 24 GB | 16.1 GB | 199 | flint | Mac M2 24GB, Mac M3 24GB, Mac M4 Pro 24GB |
| 32 GB | 21.4 GB | 202 | Kimi Linear 48B A3B Instruct | Mac M1 Pro 32GB, Mac M4 32GB |
| 36 GB | 24.1 GB | 202 | Kimi Linear 48B A3B Instruct | Mac M3 Pro 36GB |
| 48 GB | 32.2 GB | 207 | Hunyuan A13B Instruct | Mac M4 Pro 48GB |
| 64 GB | 48.0 GB | 215 | Ling 3.0 flash Sante | Mac M1 Max 64GB, Mac M4 Max 64GB, Mac M4 Pro 64GB |
| 96 GB | 72.0 GB | 219 | Step 3.5 Flash | Mac M2 Max 96GB |
| 128 GB | 96.0 GB | 227 | Intern S1 | Mac M3 Max 128GB, Mac M4 Max 128GB |
| 192 GB | 144.0 GB | 245 | Llama 4 Maverick 17B 128E Instruct | Mac M2 Ultra 192GB |
| 256 GB | 192.0 GB | 248 | MiniMax M1 40k | Mac M3 Ultra 256GB |
| 512 GB | 384.0 GB | 278 | Kimi K2.5 | Mac M3 Ultra 512GB |
Popular models that fit a 16 GB Mac
Best quality quant that fits the Mac M4 16GB at 8K context, sorted by Hugging Face downloads. Full list for this Mac.
| Model | Params | Best quant | Needs | Max tok/s (est.) |
|---|---|---|---|---|
| Qwen3 0.6B | 752M | F16 | 2.7 GB | 100 |
| Qwen3 14B | 15B | Q3_K_M | 8.7 GB | 17 |
| Qwen3 VL 8B Instruct | 8.8B | Q6_K | 8.5 GB | 17 |
| Qwen2.5 7B Instruct | 7.6B | Q8_0 | 8.7 GB | 15 |
| Qwen3.5 9B | 9.7B | Q6_K | 8.1 GB | 16 |
| Qwen3 8B | 8.2B | Q6_K | 8.1 GB | 18 |
| Unlimited OCR | 3.3B | F16 | 6.6 GB | 60 |
| Qwen2.5 VL 7B Instruct | 8.3B | Q8_0 | 8.7 GB | 15 |
| gemma 4 26B A4B it | 26B | Q2_K | 10.4 GB | 80 |
| Qwen2.5 1.5B Instruct | 1.5B | F16 | 3.8 GB | 39 |
| Qwen3.5 2B | 2.3B | F16 | 4.5 GB | 32 |
| Qwen3.5 4B | 4.7B | F16 | 9.0 GB | 14 |
| Llama 3.1 8B Instruct | 8.0B | Q8_0 | 9.6 GB | 14 |
| Qwen2.5 0.5B Instruct | 494M | F16 | 1.7 GB | 121 |
| Qwen3 VL 4B Instruct | 4.4B | F16 | 9.3 GB | 15 |
| Qwen3 4B | 4.0B | F16 | 9.3 GB | 15 |
| Qwen3 1.7B | 2.0B | F16 | 4.8 GB | 35 |
| Qwen3.5 0.8B | 873M | F16 | 2.4 GB | 79 |
| Qwen3 4B Instruct 2507 | 4.0B | F16 | 9.3 GB | 15 |
| Qwen2.5 3B Instruct | 3.1B | F16 | 6.7 GB | 19 |
Questions
Can I run an LLM on a Mac?
Yes. We track 285 open models across 18 Apple Silicon Macs. Open your exact machine above for the full list, best quant and an estimated speed ceiling.
Why does a 16 GB Mac fit fewer models than a 16 GB NVIDIA GPU?
macOS lets the GPU wire about 67% of unified memory on machines under 64 GB, and about 75% on 64 GB and larger, by default. A discrete GPU is modelled at 95% of VRAM. You can raise the Mac limit with the iogpu.wired_limit_mb setting; the numbers here stay on the defaults so they match a stock machine.
Looking for a discrete GPU instead? See models by VRAM size or the full hardware list. Methodology · Report an error.