Apple M5 Max

Laptop

Apple unified-memory SoC, launched with the 2026-03 MacBook Pro refresh: up to 128 GB of unified memory at 614 GB/s (460 GB/s on the 32-core-GPU variant). Also offered in Mac Studio. Apple does not publish a TDP, so that column is left empty.

149

Run records

38

Models covered

4

Frameworks covered

L0 Self-reported × 6 · L2 Cross-framework verified × 143

Evidence mix

Typical measured performance

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

  1. llama.cpp · IQ4_XS Decode 75.66 · Prefill 303.23 ·
  2. Ollama · IQ4_XS Decode 67.21 · Prefill 271.49 ·
  3. llama.cpp · Q3_K_S Decode 215.47 · Prefill 840.6 ·
  4. Ollama · Q2_K Decode 217.16 · Prefill 1089.96 ·
  5. llama.cpp · Q4_K_M Decode 76.62 · Prefill 265.83 ·
  6. Ollama · Q4_K_M Decode 69.91 · Prefill 321.78 ·
  7. llama.cpp · Q3_K_M Decode 152.51 · Prefill 687.68 ·
  8. Ollama · IQ4_XS Decode 123.41 · Prefill 650.98 ·
  9. Ollama · Q2_K Decode 153.39 · Prefill 803.06 ·
  10. llama.cpp · IQ2_XS Decode 103.96 · Prefill 353.6 ·
  11. llama.cpp · IQ4_XS Decode 86.99 · Prefill 341.88 ·
  12. Ollama · IQ1_S Decode 109.91 · Prefill 383.57 ·
  13. Ollama · Q5_K_S Decode 61.49 · Prefill 244.43 ·
  14. llama.cpp · IQ4_XS Decode 59.77 · Prefill 305.47 ·
  15. llama.cpp · Q3_K_M Decode 63.7 · Prefill 294.17 ·
  16. Ollama · FP16 Decode 25.66 · Prefill 97.32 ·
  17. llama.cpp · Q5_K_S Decode 59.51 · Prefill 259.7 ·
  18. Ollama · FP16 Decode 30.75 · Prefill 124.45 ·
  19. Ollama · Q5_K_S Decode 53.6 · Prefill 253.24 ·
  20. llama.cpp · Q3_K_L Decode 42.74 · Prefill 158.84 ·
  21. llama.cpp · Q6_K Decode 30.73 · Prefill 132.88 ·
  22. Ollama · IQ3_XS Decode 46.88 · Prefill 178.13 ·
  23. llama.cpp · IQ1_S Decode 83.43 · Prefill 309.76 ·
  24. Ollama · IQ4_XS Decode 48.58 · Prefill 245.06 ·
  25. llama.cpp · IQ3_XS Decode 159.32 · Prefill 706.13 ·
  26. llama.cpp · Q6_K Decode 90.4 · Prefill 328.52 ·
  27. Ollama · Q3_K_L Decode 131.28 · Prefill 555.46 ·
  28. Ollama · Q5_K_S Decode 105.03 · Prefill 472.27 ·
  29. llama.cpp · Q3_K_L Decode 104.81 · Prefill 351.35 ·
  30. Ollama · IQ1_S Decode 115.64 · Prefill 505.04 ·
  31. Ollama · Q5_K_S Decode 65.93 · Prefill 214.6 ·
  32. llama.cpp · Q2_K Decode 65.62 · Prefill 279.34 ·
  33. Ollama · IQ2_M Decode 64.33 · Prefill 209.5 ·
  34. Ollama · IQ3_XS Decode 62.24 · Prefill 282.1 ·
  35. llama.cpp · IQ3_S Decode 184.05 · Prefill 824.37 ·
  36. Ollama · IQ3_S Decode 248.21 · Prefill 1111.71 ·
  37. Ollama · IQ4_XS Decode 157.93 · Prefill 586.06 ·
  38. llama.cpp · IQ3_XS Decode 79.08 · Prefill 304.15 ·
  39. Ollama · Q2_K Decode 78.89 · Prefill 425.19 ·
  40. Ollama · Q3_K_S Decode 73.12 · Prefill 238.27 ·
  41. llama.cpp · IQ1_S Decode 95.3 · Prefill 467.44 ·
  42. Ollama · IQ2_M Decode 82.41 · Prefill 283.9 ·
  43. Ollama · Q6_K Decode 50.37 · Prefill 248.31 ·
  44. llama.cpp · IQ1_S Decode 64.71 · Prefill 267.16 ·
  45. llama.cpp · IQ2_XS Decode 58.24 · Prefill 303.37 ·
  46. Ollama · Q4_K_M Decode 32.9 · Prefill 145.14 ·
  47. llama.cpp · IQ3_S Decode 67.49 · Prefill 263.38 ·
  48. Ollama · IQ2_M Decode 66.82 · Prefill 219.12 ·
  49. Ollama · Q5_K_M Decode 43.37 · Prefill 143.55 ·
  50. llama.cpp · IQ2_M Decode 101.46 · Prefill 460.44 ·

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 Max
VRAM
128 GB
Memory bandwidth
614 GB/s

Run records

50 configurations

ModelFrameworkQuantizationDecodeGen timeSamples
Ministral-3-8B-Instruct-2512llama.cppIQ4_XS75.7 tok/s—1
Ministral-3-8B-Instruct-2512OllamaIQ4_XS67.2 tok/s—1
iFlytek Spark-X2.5-4Bllama.cppQ3_K_S215.5 tok/s—1
iFlytek Spark-X2.5-4BOllamaQ2_K217.2 tok/s—1
NVIDIA-Nemotron-3.5-Bolt-E4B-4GBllama.cppQ4_K_M76.6 tok/s—1
NVIDIA-Nemotron-3.5-Bolt-E4B-4GBOllamaQ4_K_M69.9 tok/s—1
Flint-40B-A2.5Bllama.cppQ3_K_M152.5 tok/s—1
Flint-40B-A2.5BOllamaIQ4_XS123.4 tok/s—1
Flint-40B-A2.5BOllamaQ2_K153.4 tok/s—1
Raon-1-mini-previewllama.cppIQ2_XS104 tok/s—1
Raon-1-mini-previewllama.cppIQ4_XS87 tok/s—1
Raon-1-mini-previewOllamaIQ1_S109.9 tok/s—1
Raon-1-mini-previewOllamaQ5_K_S61.5 tok/s—1
mammoth-1llama.cppIQ4_XS59.8 tok/s—1
mammoth-1llama.cppQ3_K_M63.7 tok/s—1
mammoth-1OllamaFP1625.7 tok/s—1
elephant-1llama.cppQ5_K_S59.5 tok/s—1
elephant-1OllamaFP1630.8 tok/s—1
elephant-1OllamaQ5_K_S53.6 tok/s—1
Qwen3.8-27Bllama.cppQ3_K_L42.7 tok/s—1
Qwen3.8-27Bllama.cppQ6_K30.7 tok/s—1
Qwen3.8-27BOllamaIQ3_XS46.9 tok/s—1
Qwen3.6-27Bllama.cppIQ1_S83.4 tok/s—1
Qwen3.6-27BOllamaIQ4_XS48.6 tok/s—1
Qwen3.6-35B-A3Bllama.cppIQ3_XS159.3 tok/s—1
Qwen3.6-35B-A3Bllama.cppQ6_K90.4 tok/s—1
Qwen3.6-35B-A3BOllamaQ3_K_L131.3 tok/s—1
Qwen3.6-35B-A3BOllamaQ5_K_S105 tok/s—1
Qwen3.5-4Bllama.cppQ3_K_L104.8 tok/s—1
Qwen3.5-4BOllamaIQ1_S115.6 tok/s—1
Qwen3.5-4BOllamaQ5_K_S65.9 tok/s—1
Qwen3.5-27Bllama.cppQ2_K65.6 tok/s—1
Qwen3.5-27BOllamaIQ2_M64.3 tok/s—1
Qwen3.5-27BOllamaIQ3_XS62.2 tok/s—1
Qwen3.5-0.8Bllama.cppIQ3_S184.1 tok/s—1
Qwen3.5-0.8BOllamaIQ3_S248.2 tok/s—2
Qwen3.5-0.8BOllamaIQ4_XS157.9 tok/s—1
Qwen3.8-9Bllama.cppIQ3_XS79.1 tok/s—1
Qwen3.8-9BOllamaQ2_K78.9 tok/s—1
Qwen3.8-9BOllamaQ3_K_S73.1 tok/s—1
Qwen3-8Bllama.cppIQ1_S95.3 tok/s—1
Qwen3-8BOllamaIQ2_M82.4 tok/s—1
Qwen3-8BOllamaQ6_K50.4 tok/s—1
gemma-4-31B-itllama.cppIQ1_S64.7 tok/s—1
gemma-4-31B-itllama.cppIQ2_XS58.2 tok/s—1
gemma-4-31B-itOllamaQ4_K_M32.9 tok/s—1
gemma-4-12B-itllama.cppIQ3_S67.5 tok/s—1
gemma-4-12B-itOllamaIQ2_M66.8 tok/s—1
gemma-4-12B-itOllamaQ5_K_M43.4 tok/s—1
Llama-3.1-8B-Instructllama.cppIQ2_M101.5 tok/s—1