FitLocally by GPUwerk

Google · 25.2B · MoE

Can you run Gemma 4 26B-A4B locally?

Gemma 4 26B-A4B is a mixture-of-experts model with 25.2B parameters, 3.8B active per token. At 4-bit or better it runs on 74 of the 120 hardware configurations listed here; the smallest is the RX 7900 XT (20 GB). On a 128 GB DGX Spark it runs at Q8_0, at roughly 52 tokens per second (reported).

Data as of 2 November 2026

Fastest machines for Gemma 4 26B-A4B

Every machine that runs it, fastest first. Filter by type; your machine is marked. Compare machines running it side by side

  1. 1RTX 5090Graphics card · 32 GB~195 tok/s ModelFit est.Q8_0 · 30 of 31 GB
  2. 2RTX PRO 6000 BlackwellGraphics card · 96 GB~195 tok/s ModelFit est.Q8_0 · 30 of 95 GB
  3. 32x RTX 5090Graphics card · 64 GB~184 tok/s ModelFit est.Q8_0 · 30 of 62 GB
  4. 42x RTX PRO 6000 BlackwellGraphics card · 192 GB~184 tok/s ModelFit est.Q8_0 · 30 of 190 GB
  5. 5RTX 4090Graphics card · 24 GB~168 tok/s ModelFit est.Q4_K_M · 17.4 of 23 GB
  6. 6RTX 3090 TiGraphics card · 24 GB~156 tok/s ModelFit est.Q4_K_M · 17.4 of 23 GB
  7. 7RTX 3090Graphics card · 24 GB~147 tok/s ModelFit est.Q4_K_M · 17.4 of 23 GB
  8. 8RX 7900 XTXGraphics card · 24 GB~126 tok/s ModelFit est.Q4_K_M · 17.4 of 23 GB
  9. 9M5 UltraMac · 96–512 GB · 5 variants~123 tok/s ModelFit est.Q8_0 · 30 GB
  10. 102x RTX 4090Graphics card · 48 GB~116 tok/s ModelFit est.Q8_0 · 30 of 46 GB
  11. 11RX 7900 XTGraphics card · 20 GB~111 tok/s ModelFit est.Q4_K_M · 17.4 of 19.0 GB
  12. 122x RTX 3090 TiGraphics card · 48 GB~109 tok/s ModelFit est.Q8_0 · 30 of 46 GB
  13. 132x RTX 3090Graphics card · 48 GB~102 tok/s ModelFit est.Q8_0 · 30 of 46 GB
  14. 142x RX 7900 XTXGraphics card · 48 GB~92 tok/s ModelFit est.Q8_0 · 30 of 46 GB
  15. 15M3 UltraMac · 96–512 GB · 5 variants~85 tok/s ModelFit est.Q8_0 · 30 GB
  16. 16M5 Max · 36 GB · 32cMac · 36 GB~84 tok/s ModelFit est.Q4_K_M · 17.4 of 24 GB
  17. 17M2 UltraMac · 64–192 GB · 7 variants~76 tok/s ModelFit est.Q8_0 · 30 GB
  18. 18M5 MaxMac · 48–128 GB · 3 variants~75 tok/s ModelFit est.Q8_0 · 30 GB
  19. 19M4 Max · 36 GB · 32cMac · 36 GB~71 tok/s ModelFit est.Q4_K_M · 17.4 of 24 GB
  20. 20M1 UltraMac · 64–128 GB · 4 variants~68 tok/s ModelFit est.Q8_0 · 30 GB
  21. 21M2 MaxMac · 32 GB · 2 variants~67 tok/s ModelFit est.Q4_K_M · 17.4 GB
  22. 22M4 MaxMac · 48–128 GB · 3 variants~63 tok/s ModelFit est.Q8_0 · 30 GB
  23. 23M1 MaxMac · 32 GB · 2 variants~60 tok/s ModelFit est.Q4_K_M · 17.4 GB
  24. 24M3 Max · 36 GB · 30cMac · 36 GB~55 tok/s ModelFit est.Q4_K_M · 17.4 of 24 GB
  25. 25DGX Spark (GB10)AI mini PC · 128 GB~52 tok/s reportedQ8_0 · 30 of 122 GB
  26. 26M3 MaxMac · 48–128 GB · 3 variants~49 tok/s ModelFit est.Q8_0 · 30 GB
  27. 27M2 MaxMac · 64–96 GB · 3 variants~46 tok/s ModelFit est.Q8_0 · 30 GB
  28. 28M5 ProMac · 48–64 GB · 4 variants~43 tok/s ModelFit est.Q8_0 · 30 GB
  29. 29M1 MaxMac · 64 GB · 2 variants~41 tok/s ModelFit est.Q8_0 · 30 GB
  30. 30M2 Pro · 32 GBMac · 32 GB~39 tok/s ModelFit est.Q4_K_M · 17.4 of 21 GB
  31. 31M6 · 32 GB · 12cMac · 32 GB~38 tok/s ModelFit est.Q4_K_M · 17.4 of 21 GB
  32. 32M3 Max · 96 GB · 30cMac · 96 GB~37 tok/s ModelFit est.Q8_0 · 30 of 72 GB
  33. 33M4 ProMac · 48–64 GB · 4 variants~36 tok/s ModelFit est.Q8_0 · 30 GB
  34. 34M1 Pro · 32 GBMac · 32 GB~35 tok/s ModelFit est.Q4_K_M · 17.4 of 21 GB
  35. 35Ryzen AI Max+ 395 · 64 GBAI mini PC · 64 GB~33 tok/s our est.Q8_0 · 30 of 48 GB
  36. 36M5Mac · 32 GB · 2 variants~32 tok/s ModelFit est.Q4_K_M · 17.4 GB
  37. 37M3 Pro · 36 GBMac · 36 GB~31 tok/s ModelFit est.Q4_K_M · 17.4 of 24 GB
  38. 38Ryzen AI Max+ 395 · 128 GBAI mini PC · 128 GB~29 tok/s ModelFit est.Q8_0 · 30 of 96 GB
  39. 39M4 · 32 GBMac · 32 GB~24 tok/s ModelFit est.Q4_K_M · 17.4 of 21 GB
Too big for 46 machines

M5 Pro · 24 GB · 20c, M5 Pro · 24 GB · 16c, M4 Pro · 24 GB · 20c, M4 Pro · 24 GB · 16c, M6 · 24 GB · 12c, M6 · 24 GB · 10c, M5 · 24 GB · 10c, M5 · 24 GB · 8c, M4 · 24 GB, M3 · 24 GB, M2 · 24 GB, RTX 5080, RTX 5070 Ti, RTX 5060 Ti, RTX 4080 SUPER, RTX 4080, RTX 4070 Ti SUPER, RTX 4060 Ti, Tesla T4, Tesla P100, RX 9070 XT, RX 9060 XT, M3 Pro · 18 GB, RTX 5070, RTX 4070 Ti, RTX 4070 SUPER, RTX 4070, RTX 3060, M2 Pro · 16 GB, M1 Pro · 16 GB, M6 · 16 GB · 12c, M6 · 16 GB · 10c, M5 · 16 GB · 10c, M5 · 16 GB · 8c, M4 · 16 GB, M3 · 16 GB, M2 · 16 GB, M1 · 16 GB, RTX 5060, RTX 4060, RTX 3070, RTX 3060 Ti, RTX 2070, M3 · 8 GB, M2 · 8 GB, M1 · 8 GB. Gemma 4 26B-A4B needs about 17.4 GB at its smallest build.

Speeds are single-user decode estimates, not measurements. "ModelFit est." comes from ModelFit's model fitted on published runs (CC BY 4.0); "our est." is memory bandwidth divided by bytes read per token, times 0.55. Memory is what the best build needs, against what the machine can give a model. Method

DGX Spark, reported

52 tok/s

single user. Source and setup in the GPUwerk guide

About the model

Parameters
25.2B
Active per token
3.8B
Type
Mixture of experts
Good for
Fast multimodal MoE (text and image)
Model card
Hugging Face
Hugging Face count
25.81B (agrees)

Builds and how to run them

QuantizationLoadMin. RAMCommand
Q4_K_M16.0 GB24 GBollama run gemma4:26b
Q8_028 GB48 GBollama run gemma4:26b-a4b-it-q8_0

Load, minimum RAM and Ollama commands: ModelFit, CC BY 4.0. Load is memory with the model loaded, before long contexts.

Questions

Can I run Gemma 4 26B-A4B on a Mac?

Yes on a Mac with enough memory: the smallest Apple configuration listed here that runs it at 4-bit or better is the M2 Max · 32 GB · 38c.

How much memory does Gemma 4 26B-A4B need?

About 17.4 GB at the smallest build in our data, including a margin for the runtime and an 8K-token context. Longer contexts need more.

Does Gemma 4 26B-A4B run on a DGX Spark?

Yes, at Q8_0. The Spark has 128 GB of unified memory, about 121.6 GB usable.