Gemma 4 26B (MoE)
Gemma 4 26B (MoE) needs roughly 13.0GB VRAM at Q4 quantization (52.0GB at FP16). 55 GPUs we track can run it fully in VRAM at 8k context.
Google26B params3.8B active (MoE)250k contextApache 2.0Commercial use ok
VRAM at each quantization
Assumes 8k context. KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP16 | 52.0 GB | 1.41 GB | 59.8 GB |
| Q8 | 26.0 GB | 1.41 GB | 30.7 GB |
| Q6_K | 19.5 GB | 1.41 GB | 23.4 GB |
| Q5_K_M | 16.3 GB | 1.41 GB | 19.8 GB |
| Q4_K_M | 13.0 GB | 1.41 GB | 16.1 GB |
| Q3_K_M | 10.4 GB | 1.41 GB | 13.2 GB |
| Q2_K | 7.8 GB | 1.41 GB | 10.3 GB |
Benchmarks
GPUs that run Gemma 4 26B (MoE) natively (55)
- NVIDIA RTX 5090Q6_K · 691.6 t/s
- NVIDIA RTX 4090Q5_K_M · 466.9 t/s
- NVIDIA RTX 4080Q3_K_M · 518.9 t/s
- NVIDIA RTX 4070 TiQ2_K · 486.3 t/s
- NVIDIA RTX 4070Q2_K · 486.3 t/s
- NVIDIA RTX 4060 Ti 16GBQ3_K_M · 208.4 t/s
- NVIDIA RTX 3090Q5_K_M · 433.5 t/s
- NVIDIA RTX 3090 TiQ5_K_M · 466.9 t/s
- NVIDIA RTX 3060 12GBQ2_K · 347.4 t/s
- NVIDIA H100 80GBFP16 · 484.9 t/s
- NVIDIA A100 80GBFP16 · 295.1 t/s
- NVIDIA A100 40GBQ8 · 450.1 t/s
- NVIDIA L40SQ8 · 250.1 t/s
- NVIDIA RTX A6000Q8 · 222.3 t/s
- NVIDIA RTX 6000 AdaQ8 · 277.9 t/s
- AMD Radeon RX 7900 XTXQ5_K_M · 444.6 t/s
- AMD Radeon RX 7900 XTQ4_K_M · 463.2 t/s
- AMD Radeon RX 6800 XTQ3_K_M · 370.5 t/s
- AMD Instinct MI300XFP16 · 767.1 t/s
- Apple M4 Ultra (384GB)FP16 · 158.1 t/s
- Apple M4 Ultra (192GB)FP16 · 158.1 t/s
- Apple M4 Max (128GB)FP16 · 79 t/s
- Apple M4 Max (96GB)FP16 · 79 t/s
- Apple M4 Max (64GB)FP16 · 79 t/s
- Apple M4 Max (48GB)Q8 · 158.1 t/s
- Apple M4 Pro (48GB)Q8 · 79 t/s
- Apple M4 Pro (24GB)Q5_K_M · 126.4 t/s
- Apple M4 (32GB)Q6_K · 46.3 t/s
- Apple M4 (16GB)Q2_K · 115.8 t/s
- Apple M3 Max (128GB)FP16 · 57.9 t/s
- Apple M3 Max (96GB)FP16 · 57.9 t/s
- Apple M3 Max (64GB)FP16 · 57.9 t/s
- Apple M3 Max (48GB)Q8 · 115.8 t/s
- Apple M3 Max (36GB)Q8 · 115.8 t/s
- Apple M3 Pro (36GB)Q8 · 43.4 t/s
- Apple M3 Pro (18GB)Q3_K_M · 108.6 t/s
- Apple M3 (24GB)Q5_K_M · 46.3 t/s
- Apple M3 (16GB)Q2_K · 96.5 t/s
- Apple M2 Ultra (384GB)FP16 · 115.8 t/s
- Apple M2 Ultra (192GB)FP16 · 115.8 t/s
- Apple M2 Max (96GB)FP16 · 57.9 t/s
- Apple M2 Max (64GB)FP16 · 57.9 t/s
- Apple M2 Max (32GB)Q6_K · 154.4 t/s
- Apple M2 Pro (32GB)Q6_K · 77.2 t/s
- Apple M2 Pro (16GB)Q2_K · 193 t/s
- Apple M2 (24GB)Q5_K_M · 46.3 t/s
- Apple M2 (16GB)Q2_K · 96.5 t/s
- Apple M1 Ultra (128GB)FP16 · 115.8 t/s
- Apple M1 Ultra (64GB)FP16 · 115.8 t/s
- Apple M1 Max (64GB)FP16 · 57.9 t/s
- Apple M1 Max (32GB)Q6_K · 154.4 t/s
- Apple M1 Pro (32GB)Q6_K · 77.2 t/s
- Apple M1 Pro (16GB)Q2_K · 193 t/s
- Apple M1 (16GB)Q2_K · 65.6 t/s
- Intel Arc A770 16GBQ3_K_M · 405.3 t/s
Plus 3 GPUs that run it with CPU offload (slower)
- NVIDIA RTX 4060Q8 · 17.9 t/s
- NVIDIA RTX 3080 10GBQ8 · 50 t/s
- CPU only (system RAM)Q6_K · 3.7 t/s
Notes
MoE architecture optimizing latency, activating ~3.8B parameters.