Which machines run Gemma 4 26B-A4B?
google/gemma-4-26B-A4B-it · 25.8B parameters, 4B active (MoE) · license see model card · on Hugging Face since 2026-03-11
Memory for weights (estimate)
Q4_K_M: ≈15.6 GB
Q8_0: ≈27.4 GB
Computed from parameter count × bits per weight. Plus context.
Fits at 4-bit
24 of 24 tracked devices
Measured results
12 public tok/s results
Speeds come from different engines and quants.
| Device | Memory | Q4_K_M | Q8_0 | tok/s (measured) | Quant, engine | $ per tok/s | € per tok/s |
|---|---|---|---|---|---|---|---|
| Mac Studio M5 Max 40-core GPU, 128GB / 512GB | 128GB | fits | fits | 96.3 | QAT UD-Q4_K_XL, llama.cpp, 256K context | $53 | €61 |
| MacBook Pro 16" M5 Max 40-core GPU, 128GB / 2TB | 128GB | fits | fits | 96.3 | QAT UD-Q4_K_XL, llama.cpp, 256K context | $73 | €83 |
| GMKtec EVO-X2 128GB / 1TB | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $65 | €61 |
| GMKtec EVO-X2 128GB / 2TB | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $67 | €64 |
| GMKtec EVO-X3 128GB / 2TB | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $70 | €66 |
| Minisforum MS-S1 MAX 128GB / 2TB | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $72 | €64 |
| Beelink GTR9 Pro 128GB / 2TB | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $80 | €71 |
| Framework Desktop 128GB (DIY) | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $64 | €72 |
| HP Z2 Mini G1a (Ryzen AI Max+ PRO 395) | 128GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | – | – |
| GMKtec EVO-X2 64GB / 1TB | 64GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $41 | €37 |
| Minisforum MS-S1 MAX 64GB / 2TB | 64GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $48 | €43 |
| Framework Desktop 64GB (DIY) | 64GB | fits | fits | 54.18 | UD-Q4_K_M, llama.cpp Vulkan llama-bench tg128 | $46 | €41 |
| Mac Studio M5 Ultra 80-core GPU, 256GB / 1TB | 256GB | fits | fits | no public result | – | – | |
| NVIDIA DGX Spark (Founders Edition) | 128GB | fits | fits | no public result | – | – | |
| ASUS Ascent GX10 (GB10) | 128GB | fits | fits | no public result | – | – | |
| Dell Pro Max / Pro Precision with GB10, 128GB / 2TB | 128GB | fits | fits | no public result | – | – | |
| MSI EdgeXpert MS-C931 (GB10) | 128GB | fits | fits | no public result | – | – | |
| Lenovo ThinkStation PGX (GB10) | 128GB | fits | fits | no public result | – | – | |
| Mac Studio M5 Ultra 64-core GPU, 96GB / 1TB | 96GB | fits | fits | no public result | – | – | |
| Mac mini M5 Pro 16-core GPU, 64GB / 512GB | 64GB | fits | fits | no public result | – | – | |
| Mac mini M5 Pro 20-core GPU, 64GB / 512GB | 64GB | fits | fits | no public result | – | – | |
| NVIDIA RTX 5090 32GB (new) card, needs host PC | 32GB | fits | doesn't fit | no public result | – | – | |
| NVIDIA RTX 4090 24GB (used) card, needs host PC | 24GB | fits | doesn't fit | no public result | – | – | |
| NVIDIA RTX 3090 24GB (used) card, needs host PC | 24GB | fits | doesn't fit | no public result | – | – |
Fit is an estimate: weights × 1.15 + 2 GB must fit in usable memory (VRAM − 1 GB for cards, 85% of unified memory for machines). tok/s is text generation from the linked public source for that chip; "no public result" means we found none, never a guess. Why this model is listed: MoE, 4B active; 11M downloads.