AMD Strix Halo (Radeon 8060S)

Other

47

Run records

15

Models covered

4

Frameworks covered

L2 Cross-framework verified × 47

Evidence mix

Typical measured performance

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

  1. vLLM · Q8_0 Decode 72.19 · Prefill 307.07 ·
  2. Ollama · IQ3_XS Decode 123.41 · Prefill 403.1 ·
  3. llama.cpp · IQ1_S Decode 116.24 · Prefill 433.01 ·
  4. llama.cpp · Q5_K_S Decode 60.47 · Prefill 213.4 ·
  5. Strata · Q4_K_S Decode 105.67 · Prefill 508.86 ·
  6. Strata · Q6_K Decode 73.08 · Prefill 234.71 ·
  7. Ollama · IQ3_S Decode 74.72 · Prefill 343.92 ·
  8. Strata · IQ3_XS Decode 105.76 · Prefill 504.03 ·
  9. Strata · Q5_K_M Decode 70.97 · Prefill 388.51 ·
  10. llama.cpp · Q3_K_L Decode 205.24 · Prefill 943.68 ·
  11. llama.cpp · Q3_K_M Decode 229.54 · Prefill 1241.69 ·
  12. vLLM · IQ3_S Decode 249.57 · Prefill 1015.71 ·
  13. vLLM · Q3_K_L Decode 238.07 · Prefill 894.22 ·
  14. vLLM · Q3_K_L Decode 88.34 · Prefill 283.42 ·
  15. vLLM · Q8_0 Decode 54.8 · Prefill 273.05 ·
  16. Strata · IQ3_XS Decode 138.39 · Prefill 445.74 ·
  17. Strata · Q5_K_S Decode 95.58 · Prefill 525.15 ·
  18. vLLM · IQ1_S Decode 127.84 · Prefill 540.46 ·
  19. vLLM · Q4_K_S Decode 94.74 · Prefill 425.68 ·
  20. Ollama · Q5_K_M Decode 64.59 · Prefill 334.15 ·
  21. Ollama · Q8_0 Decode 48.28 · Prefill 205.87 ·
  22. llama.cpp · Q4_K_S Decode 106.73 · Prefill 450.35 ·
  23. Strata · Q4_K_S Decode 150.62 · Prefill 773.06 ·
  24. Strata · Q6_K Decode 108.81 · Prefill 551.07 ·
  25. llama.cpp · Q2_K Decode 188.51 · Prefill 863.98 ·
  26. llama.cpp · Q3_K_L Decode 168.83 · Prefill 829.48 ·
  27. vLLM · Q3_K_L Decode 174.98 · Prefill 636.17 ·
  28. llama.cpp · Q3_K_S Decode 117.04 · Prefill 404.22 ·
  29. llama.cpp · Q5_K_M Decode 77.99 · Prefill 423.17 ·
  30. Strata · IQ2_XS Decode 206.18 · Prefill 1005.37 ·
  31. Strata · IQ3_S Decode 166.26 · Prefill 669.87 ·
  32. llama.cpp · Q8_0 Decode 46.29 · Prefill 157.5 ·
  33. Ollama · Q2_K Decode 85.26 · Prefill 425.23 ·
  34. llama.cpp · FP16 Decode 124.48 · Prefill 656.36 ·
  35. llama.cpp · Q6_K Decode 171.39 · Prefill 894.49 ·
  36. vLLM · Q5_K_S Decode 228.54 · Prefill 950.08 ·
  37. llama.cpp · IQ4_XS Decode 127.7 · Prefill 503.71 ·
  38. Ollama · Q4_K_S Decode 115.67 · Prefill 506.03 ·
  39. Ollama · Q5_K_M Decode 87.89 · Prefill 401.06 ·
  40. vLLM · IQ2_XS Decode 355.91 · Prefill 1759.4 ·
  41. vLLM · IQ3_S Decode 267.01 · Prefill 1239.81 ·
  42. Strata · IQ3_S Decode 374.22 · Prefill 1937.23 ·
  43. Ollama · FP16 Decode 70.57 · Prefill 330.84 ·
  44. Strata · Q3_K_S Decode 228.54 · Prefill 966.84 ·
  45. Strata · Q6_K Decode 157.75 · Prefill 603.58 ·
  46. llama.cpp · IQ2_M Decode 142.24 · Prefill 506.53 ·
  47. Strata · Q8_0 Decode 92.35 · Prefill 312.7 ·

Fit check

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

VRAM is not recorded for this hardware; fit cannot be computed

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

Key specs

Vendor
AMD
Memory bandwidth
256 GB/s

Run records

47 configurations

ModelFrameworkQuantizationDecodeGen timeSamples
Qwen3.5-4BvLLMQ8_072.2 tok/s—1
Qwen3.5-4BOllamaIQ3_XS123.4 tok/s—1
Qwen3.5-9Bllama.cppIQ1_S116.2 tok/s—1
Qwen3.5-9Bllama.cppQ5_K_S60.5 tok/s—1
Qwen3.5-9BStrataQ4_K_S105.7 tok/s—1
Qwen3.5-9BStrataQ6_K73.1 tok/s—1
Qwen3.5-27BOllamaIQ3_S74.7 tok/s—1
Qwen3.5-27BStrataIQ3_XS105.8 tok/s—1
Qwen3.5-27BStrataQ5_K_M71 tok/s—1
Qwen3.5-0.8Bllama.cppQ3_K_L205.2 tok/s—1
Qwen3.5-0.8Bllama.cppQ3_K_M229.5 tok/s—1
Qwen3.5-0.8BvLLMIQ3_S249.6 tok/s—1
Qwen3.5-0.8BvLLMQ3_K_L238.1 tok/s—1
Qwen3.8-9BvLLMQ3_K_L88.3 tok/s—1
Qwen3.8-9BvLLMQ8_054.8 tok/s—1
Qwen3.8-9BStrataIQ3_XS138.4 tok/s—1
Qwen3.8-9BStrataQ5_K_S95.6 tok/s—1
Qwen3-8BvLLMIQ1_S127.8 tok/s—1
Qwen3-8BvLLMQ4_K_S94.7 tok/s—1
Qwen3-8BOllamaQ5_K_M64.6 tok/s—1
Qwen3-8BOllamaQ8_048.3 tok/s—1
Qwen3-4Bllama.cppQ4_K_S106.7 tok/s—1
Qwen3-4BStrataQ4_K_S150.6 tok/s—1
Qwen3-4BStrataQ6_K108.8 tok/s—1
Qwen3-1.7Bllama.cppQ2_K188.5 tok/s—1
Qwen3-1.7Bllama.cppQ3_K_L168.8 tok/s—1
Qwen3-1.7BvLLMQ3_K_L175 tok/s—1
Qwen3-4B-Thinking-2507llama.cppQ3_K_S117 tok/s—1
Qwen3-4B-Thinking-2507llama.cppQ5_K_M78 tok/s—1
Qwen3-4B-Thinking-2507StrataIQ2_XS206.2 tok/s—1
Qwen3-4B-Thinking-2507StrataIQ3_S166.3 tok/s—1
Qwen3-Coder-Nextllama.cppQ8_046.3 tok/s—1
Qwen3-Coder-NextOllamaQ2_K85.3 tok/s—1
Qwen2.5-0.5B-Instructllama.cppFP16124.5 tok/s—1
Qwen2.5-0.5B-Instructllama.cppQ6_K171.4 tok/s—1
Qwen2.5-0.5B-InstructvLLMQ5_K_S228.5 tok/s—1
Qwen2.5-Omni-3Bllama.cppIQ4_XS127.7 tok/s—1
Qwen2.5-Omni-3BOllamaQ4_K_S115.7 tok/s—1
Qwen2.5-Omni-3BOllamaQ5_K_M87.9 tok/s—1
Qwen3-Embedding-0.6BvLLMIQ2_XS355.9 tok/s—1
Qwen3-Embedding-0.6BvLLMIQ3_S267 tok/s—1
Qwen3-Embedding-0.6BStrataIQ3_S374.2 tok/s—1
gemma-4-E2B-itOllamaFP1670.6 tok/s—1
gemma-4-E2B-itStrataQ3_K_S228.5 tok/s—1
gemma-4-E2B-itStrataQ6_K157.8 tok/s—1
gemma-4-E4B-itllama.cppIQ2_M142.2 tok/s—1
gemma-4-E4B-itStrataQ8_092.3 tok/s—1