What LLM can my Mac run? Pick your memory size
Unified memory is shared with macOS, and the GPU gets about two thirds of it on 32 GB or less and three quarters above that. The table shows the quant each Mac memory tier can use for six benchmark models at 8,192 tokens.
| Unified memory | GPU budget | Qwen3-8B | Qwen3-14B | Qwen3-30B-A3B | Qwen3-32B | Qwen2.5-72B-Instruct | Qwen3-235B-A22B | Page |
|---|---|---|---|---|---|---|---|---|
| 8 GB | 5 GiB | No | No | No | No | No | No | 8 GB Mac (2 of 22 models run) |
| 16 GB | 11 GiB | Q6_K | Q4_K_M (tight) | No | No | No | No | 16 GB Mac (8 of 22 models run) |
| 24 GB | 16 GiB | Q8_0 | Q6_K | No | No | No | No | 24 GB Mac (11 of 22 models run) |
| 32 GB | 21 GiB | Q8_0 | Q8_0 | Q4_K_M | Q4_K_M (tight) | No | No | 32 GB Mac (13 of 22 models run) |
| 48 GB | 36 GiB | Q8_0 | Q8_0 | Q8_0 | Q6_K | No | No | 48 GB Mac (17 of 22 models run) |
| 64 GB | 48 GiB | Q8_0 | Q8_0 | Q8_0 | Q8_0 | Q4_K_M (tight) | No | 64 GB Mac (17 of 22 models run) |
| 96 GB | 72 GiB | Q8_0 | Q8_0 | Q8_0 | Q8_0 | Q6_K | No | 96 GB Mac (20 of 22 models run) |
| 128 GB | 96 GiB | Q8_0 | Q8_0 | Q8_0 | Q8_0 | Q8_0 | No | 128 GB Mac (21 of 22 models run) |
llama.cpp GGUF files from the catalog, one sequence, fp16 KV cache. "No" means even Q4_K_M is over budget. The budget rule and its sources are on each tier page. These are memory estimates; speed depends on the chip's memory bandwidth, which this site does not model.