Best local LLMs for Mac by unified memory (16 to 64 GB)
On a Mac the CPU and GPU share unified memory. Here's the largest model that fits each common memory size, using the same fit rule as our calculator.
How we pick: a model counts as fitting when its Q4_K_M GGUF file plus an 8K-token FP16 KV cache uses no more than 90% of the memory, the same rule as our calculator and comparison table. File sizes are the exact bytes on Hugging Face and the KV cache comes from each model's config.json; each model page lists its sources. Ollama tags were checked on each model's Ollama library tags page and the LM Studio repo is the Hugging Face GGUF repo the size comes from (search it in LM Studio's model search). Checked 2026-11-03.
No benchmark scores here. We haven't sourced comparable coding or chat benchmark results for every model, so we don't rank on quality. Within a tier we list the biggest model that fits first, because more parameters usually means better answers, but that's a rule of thumb, not a test result.
16 GB Mac
Pick: Qwen2.5 14B Instruct, about 10.6 GB at Q4_K_M with an 8K context. Ollama: ollama run qwen2.5:14b-instruct-q4_K_M.
macOS reserves part of unified memory for itself and we have no sourced default GPU limit for 16 GB Macs, so treat these as optimistic. See the Mac guide.
| Model | Q4_K_M file | Needed at 8K context | Ollama tag | GGUF repo for LM Studio |
|---|---|---|---|---|
| Qwen2.5 14B Instruct | 9.0 GB | 10.6 GB | qwen2.5:14b-instruct-q4_K_M |
bartowski/Qwen2.5-14B-Instruct-GGUF |
| Qwen2.5-Coder 14B | 9.0 GB | 10.6 GB | qwen2.5-coder:14b-instruct-q4_K_M |
bartowski/Qwen2.5-Coder-14B-Instruct-GGUF |
| Qwen3 14B | 9.0 GB | 10.3 GB | qwen3:14b-q4_K_M |
bartowski/Qwen_Qwen3-14B-GGUF |
| Phi-4 14B | 9.1 GB | 10.7 GB | phi4:14b-q4_K_M |
bartowski/phi-4-GGUF |
| Gemma 4 12B | 7.1 GB | 7.8 GB | gemma4:12b-it-q4_K_M |
unsloth/gemma-4-12b-it-GGUF |
| Qwen3.5 9B | 6.2 GB | 6.4 GB | qwen3.5:9b-q4_K_M |
bartowski/Qwen_Qwen3.5-9B-GGUF |
8 more fit; see the 16 GB tier page for all of them.
24 GB Mac
Pick: Qwen3.8 27B, about 18.0 GB at Q4_K_M with an 8K context. Ollama: ollama run qwen3.8:27b-q4_K_M.
macOS doesn't let the GPU use all unified memory: a llama.cpp user's log shows about 17.2 GB (16,383 MiB) available to the GPU by default on a 24 GB M3. Models needing more than that may require raising the limit. Sources on the Mac guide.
| Model | Q4_K_M file | Needed at 8K context | Ollama tag | GGUF repo for LM Studio |
|---|---|---|---|---|
| Qwen3.8 27B | 17.4 GB | 18.0 GB | qwen3.8:27b-q4_K_M |
bartowski/Qwen3.8-27B-GGUF |
| Devstral Small 2 24B | 14.3 GB | 15.7 GB | devstral-small-2:24b-instruct-2512-q4_K_M |
bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF |
| Qwen2.5 14B Instruct | 9.0 GB | 10.6 GB | qwen2.5:14b-instruct-q4_K_M |
bartowski/Qwen2.5-14B-Instruct-GGUF |
| Qwen2.5-Coder 14B | 9.0 GB | 10.6 GB | qwen2.5-coder:14b-instruct-q4_K_M |
bartowski/Qwen2.5-Coder-14B-Instruct-GGUF |
| Qwen3 14B | 9.0 GB | 10.3 GB | qwen3:14b-q4_K_M |
bartowski/Qwen_Qwen3-14B-GGUF |
| Phi-4 14B | 9.1 GB | 10.7 GB | phi4:14b-q4_K_M |
bartowski/phi-4-GGUF |
10 more fit; see the 24 GB tier page for all of them.
32 GB Mac
Pick: Qwen3.6 35B-A3B, about 22.5 GB at Q4_K_M with an 8K context. Ollama: ollama run qwen3.6:35b-a3b-q4_K_M.
macOS doesn't let the GPU use all unified memory: Apple's example gives the GPU 21 GB by default on a 32 GB M1 Pro/Max MacBook Pro. Models needing more than that may require raising the limit. Sources on the Mac guide.
| Model | Q4_K_M file | Needed at 8K context | Ollama tag | GGUF repo for LM Studio |
|---|---|---|---|---|
| Qwen3.6 35B-A3B | 22.3 GB | 22.5 GB | qwen3.6:35b-a3b-q4_K_M |
bartowski/Qwen_Qwen3.6-35B-A3B-GGUF |
| Qwen2.5 32B Instruct | 19.9 GB | 22.0 GB | qwen2.5:32b-instruct-q4_K_M |
bartowski/Qwen2.5-32B-Instruct-GGUF |
| Qwen2.5-Coder 32B | 19.9 GB | 22.0 GB | qwen2.5-coder:32b-instruct-q4_K_M |
bartowski/Qwen2.5-Coder-32B-Instruct-GGUF |
| Qwen3 32B | 19.8 GB | 21.9 GB | qwen3:32b-q4_K_M |
bartowski/Qwen_Qwen3-32B-GGUF |
| Qwen3.8 27B | 17.4 GB | 18.0 GB | qwen3.8:27b-q4_K_M |
bartowski/Qwen3.8-27B-GGUF |
| Devstral Small 2 24B | 14.3 GB | 15.7 GB | devstral-small-2:24b-instruct-2512-q4_K_M |
bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF |
14 more fit; see the 32 GB tier page for all of them.
64 GB Mac
Pick: Llama 3.1 70B Instruct, about 45.2 GB at Q4_K_M with an 8K context. Ollama: ollama run llama3.1:70b-instruct-q4_K_M.
macOS doesn't let the GPU use all unified memory: Apple's example gives the GPU 48 GB by default on a 64 GB M1 Pro/Max MacBook Pro. Models needing more than that may require raising the limit. Sources on the Mac guide.
| Model | Q4_K_M file | Needed at 8K context | Ollama tag | GGUF repo for LM Studio |
|---|---|---|---|---|
| Llama 3.1 70B Instruct | 42.5 GB | 45.2 GB | llama3.1:70b-instruct-q4_K_M |
bartowski/Meta-Llama-3.1-70B-Instruct-GGUF |
| Llama 3.3 70B Instruct | 42.5 GB | 45.2 GB | llama3.3:70b-instruct-q4_K_M |
bartowski/Llama-3.3-70B-Instruct-GGUF |
| Qwen3.6 35B-A3B | 22.3 GB | 22.5 GB | qwen3.6:35b-a3b-q4_K_M |
bartowski/Qwen_Qwen3.6-35B-A3B-GGUF |
| Qwen2.5 32B Instruct | 19.9 GB | 22.0 GB | qwen2.5:32b-instruct-q4_K_M |
bartowski/Qwen2.5-32B-Instruct-GGUF |
| Qwen2.5-Coder 32B | 19.9 GB | 22.0 GB | qwen2.5-coder:32b-instruct-q4_K_M |
bartowski/Qwen2.5-Coder-32B-Instruct-GGUF |
| Qwen3 32B | 19.8 GB | 21.9 GB | qwen3:32b-q4_K_M |
bartowski/Qwen_Qwen3-32B-GGUF |
16 more fit; see the 64 GB tier page for all of them.