Gemma 4 26B-A4B — VRAM requirements

Weights = params × bytes-per-parameter for the selected quantization.KV cache is always computed at FP16 and assumes batch size 1. A flat +15% is added on top for CUDA context/activation overhead — a rule of thumb, not a measured figure. Full methodology on the interactive calculator.

Sparse mixture-of-experts. **Both numbers in the model's name are rounded up from what Google's own spec table states**: that table gives 25.2B total and 3.8B active (8 active of 128 experts, plus 1 shared), not 26B/4B. paramsB here is the exact safetensors total (25.81B), which exceeds the card's 25.2B because it also includes the ~550M vision encoder; activeParamsB uses Google's stated 3.8B. The weights term uses the total, since all 128 experts must be VRAM-resident. maxContextLength 262,144 per the card. Pulled 2026-09-19.