The best local LLM for a 24GB Mac
M4 Pro 24GB, M4 24GB, M2 24GB — about 20.4 GiB usable once macOS has taken its share.
Unified memory is the whole game on Apple Silicon. Your 24GB is shared between macOS, your apps and the model, so the honest budget is closer to 20.4 GiB than to 24. Everything below is sized against that number, at Q4_K_M with an 8K context.
The picks
Best all-round: Mistral Small 3 24B
The largest general-purpose model that still leaves room to work. It loads in 16.2 GiB and generates around 15 tokens/sec on an M4 Pro 24GB. Full breakdown →
Best for coding: Qwen2.5-Coder 14B
Trained specifically on code, and worth the swap if that is your workload. It loads in 11.2 GiB and generates around 24 tokens/sec on an M4 Pro 24GB. Full breakdown →
Best for reasoning: DeepSeek-R1-Distill-Qwen-14B
Thinks before answering; slower per question, better on hard ones. It loads in 11.2 GiB and generates around 24 tokens/sec on an M4 Pro 24GB. Full breakdown →
Fastest usable: Llama 3.2 1B
When latency matters more than depth — voice assistants, autocomplete, agents. It loads in 1.8 GiB and generates around 288 tokens/sec on an M4 Pro 24GB. Full breakdown →
Everything that fits in 24GB
| Model | Params | Loaded | Tok/s | Max ctx |
|---|---|---|---|---|
| Mistral Small 3 24B | 23.6B | 16.2 GiB | 15 | 16K |
| gpt-oss-20b | 20.9B | 13.7 GiB | 62 | 64K |
| DeepSeek-R1-Distill-Qwen-14B | 14.8B | 11.2 GiB | 24 | 32K |
| Qwen2.5 14B | 14.8B | 11.2 GiB | 24 | 32K |
| Qwen2.5-Coder 14B | 14.8B | 11.2 GiB | 24 | 32K |
| Qwen3 14B | 14.8B | 10.9 GiB | 24 | 32K |
| Phi-4 14B | 14.7B | 11.2 GiB | 24 | 16K |
| Gemma 3 12B | 12.2B | 11.1 GiB | 29 | 16K |
| Gemma 2 9B | 9.24B | 9.0 GiB | 39 | 8K |
| Qwen3 8B | 8.2B | 6.8 GiB | 43 | 64K |
| Llama 3.1 8B | 8.03B | 6.6 GiB | 44 | 64K |
| DeepSeek-R1-Distill-Qwen-7B | 7.62B | 5.8 GiB | 47 | 128K |
| Qwen2.5 7B | 7.62B | 5.8 GiB | 47 | 128K |
| Qwen2.5-Coder 7B | 7.62B | 5.8 GiB | 47 | 128K |
| Mistral 7B v0.3 | 7.25B | 6.1 GiB | 49 | 32K |
| Gemma 3 4B | 4.3B | 4.4 GiB | 83 | 64K |
| Llama 3.2 3B | 3.21B | 3.6 GiB | 111 | 64K |
| Qwen2.5 3B | 3.09B | 2.9 GiB | 115 | 32K |
| Gemma 2 2B | 2.61B | 3.0 GiB | 137 | 8K |
| Qwen2.5 1.5B | 1.54B | 1.9 GiB | 232 | 32K |
| Llama 3.2 1B | 1.24B | 1.8 GiB | 288 | 128K |
What does not fit
| Model | Needs | Short by |
|---|---|---|
| Gemma 3 27B | 21.1 GiB | 0.7 GiB |
| DeepSeek-R1-Distill-Qwen-32B | 22.4 GiB | 2.0 GiB |
| Qwen2.5 32B | 22.4 GiB | 2.0 GiB |
| Qwen2.5-Coder 32B | 22.4 GiB | 2.0 GiB |
| Qwen3 32B | 22.4 GiB | 2.0 GiB |
| Mixtral 8x7B | 29.8 GiB | 9.4 GiB |
| DeepSeek-R1-Distill-Llama-70B | 45.6 GiB | 25.2 GiB |
| Llama 3.3 70B | 45.6 GiB | 25.2 GiB |
| Qwen2.5 72B | 46.8 GiB | 26.4 GiB |
| gpt-oss-120b | 71.3 GiB | 50.9 GiB |
A model that is a gigabyte or two over can often be rescued by dropping to Q3_K_M or quantising the KV cache. Anything further over than that is better solved by picking a smaller model — a 14B at Q4 beats a 32B at Q2 on almost every task.