Gemma 4 26B-A4B (MoE): VRAM requirements and GPUs that run it

Memory needed by quantization

Totals include about 10% runtime overhead and the KV cache, plus around 0.6 GB for the vision encoder. Longer conversations cost real memory: going from 8K to 32K context adds about 4.7 GB at any quantization. It is a mixture-of-experts model: 25.8B parameters in total but only about 4B are used for each token, so it runs much faster than a dense model of the same size and stays usable even when part of it sits in system RAM. Most people should download Q4_K_M; Q5 or Q8 only make sense if you have spare memory. Read which quantization to pick.

Which GPUs can run Gemma 4 26B-A4B (MoE)?

4 of the 40 cards we track run it fully in VRAM at 8K context. The cheapest entry point is a 24 GB card such as the RTX 3090 / 3090 Ti, and the fastest is the RTX 5090 at about 150 tokens per second. 36 more can run it with part of the model in system RAM, which is noticeably slower.

Running it on a Mac

Apple Silicon Macs can give a model about 72% of their unified memory. Gemma 4 26B-A4B (MoE) fits from 24 GB (Apple M3, about 14 tokens per second). With a faster chip like the Apple M3 Ultra (96 GB) expect around 54 tokens per second. Compare the options in our Mac vs NVIDIA guide.

Similar size models: Phi-4 14B, Qwen3 14B, DeepSeek-R1 Distill 14B, gpt-oss 20B, Codestral 22B.

Compare Gemma 4 26B-A4B (MoE)

Questions

Estimates based on model size, architecture and memory bandwidth. Real numbers vary with software, drivers and context length.