Best local LLM for your Mac.
On a Mac the memory is shared with the graphics chip, so about 75% of it can hold a model. More memory = bigger models. More bandwidth = faster answers.
Published speeds.
Other people's measurements, with sources. Not our tests.
| Machine | Model | Tokens/sec | Engine | Source |
|---|---|---|---|---|
| MacBook Pro 14" (M5) | gpt-oss-20b | 48.7 | oMLX v0.3.8 · MXFP4 (MLX) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Pro) | gpt-oss-20b | 76.6 | oMLX v0.6.2 · MXFP4-Q8 (MLX) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Max) | gpt-oss-120b | 79.1 | Ollama 0.34.1 · MXFP4 (GGUF) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Max) | gpt-oss-20b | 113.4 | Ollama 0.34.1 · MXFP4 (GGUF) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Max) | Qwen3.8 27B | 36.6 | Ollama 0.34.1 · GGUF Q4 (17 GB) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Max) | Qwen3.8 27B | 68 | Ollama 0.34.1 (MLX) · MLX (18 GB) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Max) | Qwen3.6 35B-A3B | 167.2 | Ollama 0.34.1 (MLX) · MLX (23 GB) | 2026-10-15 |
| MacBook Pro 14"/16" (M5 Max) | Gemma 4 12B | 58.1 | Ollama 0.34.1 · GGUF Q4 (7.6 GB) | 2026-10-15 |
| Mac mini (M5 Pro) | gpt-oss-20b | 76.6 | oMLX v0.6.2 · MXFP4-Q8 (MLX) | 2026-10-15 |
| Mac Studio (M5 Max) | gpt-oss-120b | 78.2 | oMLX v0.3.9 · MXFP4-Q8 (MLX) | 2026-10-15 |
| Mac Studio (M5 Max) | gpt-oss-120b | 70.5 | oMLX v0.3.9 · MXFP4-Q8 (MLX) | 2026-10-15 |
| Mac Studio (M5 Ultra) | gpt-oss-120b | 108.6 | oMLX v0.7.0 · MXFP4 (MLX) | 2026-10-15 |
| Mac Studio (M5 Ultra) | gpt-oss-120b | 110.1 | oMLX v0.7.0 · MXFP4 (MLX) | 2026-10-15 |
Short answers.
- What is the best local LLM for the smallest Mac?
- With 16 GB (MacBook Air 13"/15" (M5)), the best fit is K2 Horizon 7B (needs 8.6 GB). 38 models fit.
- Which Mac runs the smartest local model?
- Mac Studio (M5 Ultra) with 512 GB runs GLM-5.3-Flash (AA 41.8, about 73% of Claude Opus 5.5).
- What matters more, memory or speed?
- Memory decides which models fit at all. Bandwidth (GB/s) decides how fast they answer. Pick memory first, then the fastest machine you can afford with that memory.
Scores are from Artificial Analysis and LMArena, not our own tests yet. Memory need = 4-bit file × 1.15 + 2 GB; a model fits when that is at most 75% of memory. Models 2026-10-15, machines 2026-10-15.
# Best local LLM for your Mac (MacBook, Mac mini, Mac Studio) URL: https://insidethemachine.ai/best/local-llm-for-mac Data: machines 2026-10-15, models 2026-10-15. On a Mac the memory is shared with the graphics chip, so about 75% of it can hold a model. More memory = bigger models. More bandwidth = faster answers. | machine | memory GB | usable GB | bandwidth GB/s | models that fit | best fit | needs GB | |---|---|---|---|---|---|---| | MacBook Air 13"/15" (M5) | 16 | 12 | 153 | 38 | K2 Horizon 7B | 8.6 | | MacBook Air 13"/15" (M5) | 24 | 18 | 153 | 45 | K2 Horizon 7B | 8.6 | | MacBook Air 13"/15" (M5) | 32 | 24 | 153 | 62 | Qwen3.8 27B | 22.7 | | MacBook Pro 14" (M5) | 16 | 12 | 153 | 38 | K2 Horizon 7B | 8.6 | | MacBook Pro 14" (M5) | 24 | 18 | 153 | 45 | K2 Horizon 7B | 8.6 | | MacBook Pro 14" (M5) | 32 | 24 | 153 | 62 | Qwen3.8 27B | 22.7 | | MacBook Pro 14"/16" (M5 Pro) | 24 | 18 | 307 | 45 | K2 Horizon 7B | 8.6 | | MacBook Pro 14"/16" (M5 Pro) | 48 | 36 | 307 | 76 | Qwen3.8 27B | 22.7 | | MacBook Pro 14"/16" (M5 Pro) | 64 | 48 | 307 | 77 | Qwen3.8 27B | 22.7 | | MacBook Pro 14"/16" (M5 Max) | 36 | 27 | 614 | 67 | Qwen3.8 27B | 22.7 | | MacBook Pro 14"/16" (M5 Max) | 48 | 36 | 614 | 76 | Qwen3.8 27B | 22.7 | | MacBook Pro 14"/16" (M5 Max) | 64 | 48 | 614 | 77 | Qwen3.8 27B | 22.7 | | MacBook Pro 14"/16" (M5 Max) | 128 | 96 | 614 | 90 | Qwen3.8 27B | 22.7 | | Mac mini (M6) | 16 | 12 | 153 | 38 | K2 Horizon 7B | 8.6 | | Mac mini (M6) | 24 | 18 | 153 | 45 | K2 Horizon 7B | 8.6 | | Mac mini (M6) | 32 | 24 | 153 | 62 | Qwen3.8 27B | 22.7 | | Mac mini (M5 Pro) | 24 | 18 | 307 | 45 | K2 Horizon 7B | 8.6 | | Mac mini (M5 Pro) | 48 | 36 | 307 | 76 | Qwen3.8 27B | 22.7 | | Mac mini (M5 Pro) | 64 | 48 | 307 | 77 | Qwen3.8 27B | 22.7 | | Mac Studio (M5 Max) | 36 | 27 | 614 | 67 | Qwen3.8 27B | 22.7 | | Mac Studio (M5 Max) | 48 | 36 | 614 | 76 | Qwen3.8 27B | 22.7 | | Mac Studio (M5 Max) | 64 | 48 | 614 | 77 | Qwen3.8 27B | 22.7 | | Mac Studio (M5 Max) | 128 | 96 | 614 | 90 | Qwen3.8 27B | 22.7 | | Mac Studio (M5 Ultra) | 96 | 72 | 1200 | 83 | Qwen3.8 27B | 22.7 | | Mac Studio (M5 Ultra) | 256 | 192 | 1200 | 95 | Qwen3.8-Flash-Next (125B-A6B) | 140 | | Mac Studio (M5 Ultra) | 512 | 384 | 1200 | 102 | GLM-5.3-Flash | 233 | ## Published speeds (decode tokens/sec, not our tests) - MacBook Pro 14" (M5): gpt-oss-20b 48.7 tok/s (oMLX v0.3.8, MXFP4 (MLX), M5 10-core GPU, 32GB) source: https://omlx.ai/benchmarks/performance/xns0rgnb (2026-10-15) - MacBook Pro 14"/16" (M5 Pro): gpt-oss-20b 76.6 tok/s (oMLX v0.6.2, MXFP4-Q8 (MLX), M5 Pro 20-core GPU, 48GB) source: https://omlx.ai/benchmarks/performance/h03e0nvx (2026-10-15) - MacBook Pro 14"/16" (M5 Max): gpt-oss-120b 79.1 tok/s (Ollama 0.34.1, MXFP4 (GGUF), 16" MBP M5 Max 40c, 128GB (Joel Ramat via bestllmfor.com, 2026-09-16)) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - MacBook Pro 14"/16" (M5 Max): gpt-oss-20b 113.4 tok/s (Ollama 0.34.1, MXFP4 (GGUF), 16" MBP M5 Max 40c, 128GB) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - MacBook Pro 14"/16" (M5 Max): Qwen3.8 27B 36.6 tok/s (Ollama 0.34.1, GGUF Q4 (17 GB), 16" MBP M5 Max 40c, 128GB) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - MacBook Pro 14"/16" (M5 Max): Qwen3.8 27B 68 tok/s (Ollama 0.34.1 (MLX), MLX (18 GB), 16" MBP M5 Max 40c, 128GB) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - MacBook Pro 14"/16" (M5 Max): Qwen3.6 35B-A3B 167.2 tok/s (Ollama 0.34.1 (MLX), MLX (23 GB), 16" MBP M5 Max 40c, 128GB) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - MacBook Pro 14"/16" (M5 Max): Gemma 4 12B 58.1 tok/s (Ollama 0.34.1, GGUF Q4 (7.6 GB), 16" MBP M5 Max 40c, 128GB) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - Mac mini (M5 Pro): gpt-oss-20b 76.6 tok/s (oMLX v0.6.2, MXFP4-Q8 (MLX), M5 Pro 20-core GPU, 48GB (chip-level result; device not stated)) source: https://omlx.ai/benchmarks/performance/h03e0nvx (2026-10-15) - Mac Studio (M5 Max): gpt-oss-120b 78.2 tok/s (oMLX v0.3.9, MXFP4-Q8 (MLX), M5 Max 40c, 128GB) source: https://omlx.ai/benchmarks/dgadv0k9 (2026-10-15) - Mac Studio (M5 Max): gpt-oss-120b 70.5 tok/s (oMLX v0.3.9, MXFP4-Q8 (MLX), M5 Max 40c, 128GB) source: https://omlx.ai/benchmarks/dgadv0k9 (2026-10-15) - Mac Studio (M5 Ultra): gpt-oss-120b 108.6 tok/s (oMLX v0.7.0, MXFP4 (MLX), M5 Ultra 80c, 256GB) source: https://omlx.ai/benchmarks/performance/bcoqxlcv (2026-10-15) - Mac Studio (M5 Ultra): gpt-oss-120b 110.1 tok/s (oMLX v0.7.0, MXFP4 (MLX), M5 Ultra 80c, 256GB) source: https://omlx.ai/benchmarks/performance/bcoqxlcv (2026-10-15) ## Questions **What is the best local LLM for the smallest Mac?** With 16 GB (MacBook Air 13"/15" (M5)), the best fit is K2 Horizon 7B (needs 8.6 GB). 38 models fit. **Which Mac runs the smartest local model?** Mac Studio (M5 Ultra) with 512 GB runs GLM-5.3-Flash (AA 41.8, about 73% of Claude Opus 5.5). **What matters more, memory or speed?** Memory decides which models fit at all. Bandwidth (GB/s) decides how fast they answer. Pick memory first, then the fastest machine you can afford with that memory.