Apple M5 Pro

Laptop

Apple unified-memory SoC, launched with the 2026-03 MacBook Pro refresh: up to 64 GB of unified memory at 307 GB/s, built on Apple's new Fusion architecture. Apple does not publish a TDP, so that column is left empty.

43

Run records

18

Models covered

4

Frameworks covered

L0 Self-reported × 7 · L2 Cross-framework verified × 36

Evidence mix

Typical measured performance

Published benchmark averages per model × framework × quantization on this hardware.

  1. Ollama · Q5_K_M Decode 121.74 · Prefill 552.05 ·
  2. Ollama · Q6_K Decode 111.41 · Prefill 526.06 ·
  3. vLLM · Q3_K_L Decode 145.87 · Prefill 596.68 ·
  4. vLLM · Q3_K_S Decode 149.44 · Prefill 576.02 ·
  5. llama.cpp · IQ1_S Decode 341.6 · Prefill 1406.03 ·
  6. vLLM · IQ2_M Decode 323.11 · Prefill 1474.14 ·
  7. vLLM · Q3_K_M Decode 358.52 · Prefill 1906.72 ·
  8. vLLM · IQ3_S Decode 158.08 · Prefill 657.93 ·
  9. vLLM · Q2_K Decode 169.33 · Prefill 690.81 ·
  10. Ollama · IQ3_S Decode 136.51 · Prefill 597.98 ·
  11. Ollama · Q3_K_S Decode 144.08 · Prefill 671.77 ·
  12. Ollama · IQ1_S Decode 187.29 · Prefill 770.8 ·
  13. Ollama · Q4_K_S Decode 116.66 · Prefill 579.89 ·
  14. Strata · Q3_K_L Decode 212.69 · Prefill 1110.56 ·
  15. llama.cpp · Q3_K_L Decode 189.13 · Prefill 672.33 ·
  16. llama.cpp · Q3_K_M Decode 187.03 · Prefill 648.58 ·
  17. llama.cpp · IQ2_M Decode 311.55 · Prefill 1641.2 ·
  18. llama.cpp · Q5_K_S Decode 197.87 · Prefill 839.49 ·
  19. vLLM · Q4_K_M Decode 233.37 · Prefill 1190.9 ·
  20. vLLM · Q5_K_S Decode 229.77 · Prefill 936.19 ·
  21. llama.cpp · Q3_K_L Decode 181 · Prefill 762.88 ·
  22. vLLM · Q5_K_S Decode 376.55 · Prefill 1237.86 ·
  23. vLLM · Q8_0 Decode 284.14 · Prefill 1555.29 ·
  24. Ollama · Q5_K_M Decode 149.09 · Prefill 749.3 ·
  25. llama.cpp · IQ1_S Decode 286.68 · Prefill 1382.17 ·
  26. llama.cpp · IQ2_M Decode 301.75 · Prefill 1278.13 ·
  27. Ollama · Q4_K_S Decode 342.8 · Prefill 1259.06 ·
  28. Ollama · Q3_K_M Decode 62.31 · Prefill 340.97 ·
  29. Strata · IQ3_XS Decode 111.36 · Prefill 546.98 ·
  30. llama.cpp · Q8_0 Decode 55.26 · Prefill 261.28 ·
  31. Strata · Q6_K Decode 98.9 · Prefill 379.8 ·
  32. llama.cpp · Q4_K_M Decode 100.37 · Prefill 374.04 ·
  33. llama.cpp · Q8_0 Decode 66.87 · Prefill 301.78 ·
  34. Ollama · Q3_K_M Decode 110.36 · Prefill 569.49 ·
  35. Ollama · Q3_K_S Decode 111.33 · Prefill 497.48 ·
  36. llama.cpp · Q2_K Decode 127.36 · Prefill 409.89 ·
  37. llama.cpp · IQ2_XS Decode 0.9 · Prefill 6.8 ·
  38. llama.cpp · IQ3_S Decode 0.8 · Prefill 6 ·
  39. llama.cpp · IQ2_XS Decode 0.9 · Prefill 6.8 ·
  40. llama.cpp · IQ3_S Decode 0.8 · Prefill 6 ·
  41. llama.cpp · IQ2_XS Decode 26.3 · Prefill 197.3 ·
  42. llama.cpp · Q5_K_M Decode 12 · Prefill 90 ·
  43. llama.cpp · Q4_K_M Decode 14.6 · Prefill 109.5 ·

Fit check

Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed

Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers

Key specs

Vendor
Apple
Architecture
M5 Pro
VRAM
64 GB
Memory bandwidth
307 GB/s

Run records

43 configurations

ModelFrameworkQuantizationDecodeGen timeSamples
Qwen3.5-4BOllamaQ5_K_M121.7 tok/s—1
Qwen3.5-4BOllamaQ6_K111.4 tok/s—1
Qwen3.5-9BvLLMQ3_K_L145.9 tok/s—1
Qwen3.5-9BvLLMQ3_K_S149.4 tok/s—1
Qwen3.5-0.8Bllama.cppIQ1_S341.6 tok/s—1
Qwen3.5-0.8BvLLMIQ2_M323.1 tok/s—1
Qwen3.5-0.8BvLLMQ3_K_M358.5 tok/s—1
Qwen3.8-9BvLLMIQ3_S158.1 tok/s—1
Qwen3.8-9BvLLMQ2_K169.3 tok/s—1
Qwen3.8-9BOllamaIQ3_S136.5 tok/s—1
Qwen3.8-9BOllamaQ3_K_S144.1 tok/s—1
Qwen3-8BOllamaIQ1_S187.3 tok/s—1
Qwen3-8BOllamaQ4_K_S116.7 tok/s—1
Qwen3-8BStrataQ3_K_L212.7 tok/s—1
Qwen3-4Bllama.cppQ3_K_L189.1 tok/s—1
Qwen3-4Bllama.cppQ3_K_M187 tok/s—1
Qwen3-1.7Bllama.cppIQ2_M311.6 tok/s—1
Qwen3-1.7Bllama.cppQ5_K_S197.9 tok/s—1
Qwen3-1.7BvLLMQ4_K_M233.4 tok/s—1
Qwen3-1.7BvLLMQ5_K_S229.8 tok/s—1
Qwen3-4B-Thinking-2507llama.cppQ3_K_L181 tok/s—1
Qwen2.5-0.5B-InstructvLLMQ5_K_S376.6 tok/s—1
Qwen2.5-0.5B-InstructvLLMQ8_0284.1 tok/s—1
Qwen2.5-Omni-3BOllamaQ5_K_M149.1 tok/s—1
Qwen3-Embedding-0.6Bllama.cppIQ1_S286.7 tok/s—1
Qwen3-Embedding-0.6Bllama.cppIQ2_M301.8 tok/s—1
Qwen3-Embedding-0.6BOllamaQ4_K_S342.8 tok/s—1
gemma-4-31B-itOllamaQ3_K_M62.3 tok/s—1
gemma-4-31B-itStrataIQ3_XS111.4 tok/s—1
gemma-4-26B-A4B-itllama.cppQ8_055.3 tok/s—1
gemma-4-26B-A4B-itStrataQ6_K98.9 tok/s—1
gemma-4-12B-itllama.cppQ4_K_M100.4 tok/s—1
gemma-4-12B-itllama.cppQ8_066.9 tok/s—1
gemma-4-12B-itOllamaQ3_K_M110.4 tok/s—1
gemma-4-12B-itOllamaQ3_K_S111.3 tok/s—1
DeepSeek-R1-Distill-Qwen-14Bllama.cppQ2_K127.4 tok/s—1
GLM-5.3llama.cppIQ2_XS0.9 tok/s—1
GLM-5.3llama.cppIQ3_S0.8 tok/s—1
DeepSeek-V4.1-Flashllama.cppIQ2_XS0.9 tok/s—1
DeepSeek-V4.1-Flashllama.cppIQ3_S0.8 tok/s—1
Qwen3.8-27Bllama.cppIQ2_XS26.3 tok/s—1
Qwen3.8-27Bllama.cppQ5_K_M12 tok/s—1
Qwen3.8-27Bllama.cppQ4_K_M14.6 tok/s—1