How Much VRAM Does Gemma 4 26B A4B Need to Run Locally?
See devices that run itGemma 4 26B A4B is a MoE model with 25.8B total parameters and 4B active per token. At Q4 it needs 18GB of VRAM to sit entirely on the GPU; with experts offloaded to system RAM it needs only 6GB of VRAM plus 13GB of RAM. The cheapest device that holds it is the Radeon RX 7900 XT, with no used-price data yet, at roughly 83 tokens/s. In offload mode the cheapest card that runs it is the GeForce RTX 2080 Ti, at roughly 38 tokens/s.
Gemma 4 26B A4B details
VRAM needed for Gemma 4 26B A4B by quantization and context
Total at 8KMoE: all 25.8B parameters must stay resident in memory, while speed is set by the 4B active per token.
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| Q4_K_M | 16.9GGUF | 17.3 | 22.9 | 45.4 |
| Q5_K_M | 21.2GGUF | 20.3 | 25.9 | 48.4 |
| Q8_0 | 26.9GGUF | 29.6 | 35.3 | 57.8 |
| FP16 | 50.5GGUF | 53.4 | 59 | 81.5 |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 256K limit show a dash.
Which GPUs can run Gemma 4 26B A4B?
Q4_K_M · 8K context · sorted by eBay used price| Device | VRAMGB | Used price | Est. t/s | Suggested context |
|---|---|---|---|---|
 Radeon RX 7900 XT | 20 | —No listings | 83 | 8K |
 Arc Pro B60 | 24 | —No listings | 50.8 | 32K |
 GeForce RTX 3090 | 24 | —No listings | 112.2 | 32K |
| GeForce RTX 4090 | 24 | —No listings | 116.8 | 32K |
| GeForce RTX 5090 D V2 | 24 | —No listings | 134.8 | 32K |
| Radeon RX 7900 XTX | 24 | —No listings | 93.4 | 32K |
| RTX PRO 4000 Blackwell | 24 | —No listings | 92.3 | 32K |
| Arc Pro B65 | 32 | —No listings | 63.5 | 32K |
| Arc Pro B70 | 32 | —No listings | 63.5 | 32K |
| GeForce RTX 4080 SUPER 32GB (modded) | 32 | —No listings | 97.6 | 32K |
| GeForce RTX 5090 | 32 | —No listings | 152.3 | 32K |
| GeForce RTX 5090 D | 32 | —No listings | 152.3 | 32K |
| Instinct MI100 | 32 | —No listings | 108.3 | 32K |
| Radeon AI PRO R9700 | 32 | —No listings | 71.2 | 32K |
| Radeon PRO W7800 | 32 | —No listings | 65.9 | 32K |
| RTX PRO 4500 Blackwell | 32 | —No listings | 109.5 | 32K |
| Tesla V100 32GB | 32 | —No listings | 109.8 | 32K |
| A100 40GB PCIe | 40 | —No listings | 143.8 | 32K |
| CMP 170HX 40GB (modded) | 40 | —No listings | 144 | 32K |
| GeForce RTX 4090 48GB (modded) | 48 | —No listings | 116.8 | 128K |
| Radeon PRO W7900 | 48 | —No listings | 87.4 | 128K |
| RTX PRO 5000 Blackwell 48GB | 48 | —No listings | 134.8 | 128K |
| CMP 170HX 64GB (modded) | 64 | —No listings | 141.3 | 128K |
| Instinct MI210 | 64 | —No listings | 126.1 | 128K |
| RTX PRO 5000 Blackwell 72GB | 72 | —No listings | 134.8 | 128K |
| RTX PRO 6000D | 84 | —No listings | 144.3 | 128K |
| RTX PRO 6000 Blackwell | 96 | —No listings | 152.3 | 128K |
| Mac Studio M4 Max 128GB | 128 | —No listings | 68 | 128K |
| Mac Studio M5 Max 128GB | 128 | —No listings | 73.9 | 128K |
| MacBook Pro M4 Max 128GB | 128 | —No listings | 68 | 128K |
| MacBook Pro M5 Max 128GB | 128 | —No listings | 73.9 | 128K |
| Mac Studio M3 Ultra 512GB | 512 | —No listings | 89.7 | 128K |
Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.
GPUs that run Gemma 4 26B A4B with experts offloaded to RAM
Expert weights live in system RAM (needs at least 13 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 6 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.
| Device | VRAMGB | Used price | RAM neededGB | Est. t/s |
|---|---|---|---|---|
| GeForce RTX 2080 Ti | 11 | — | 13 | 37.8 |
| GeForce RTX 4080 | 16 | — | 13 | 38.9 |
| GeForce RTX 5070 Ti | 16 | — | 13 | 40.4 |
| GeForce RTX 5080 | 16 | — | 13 | 40.9 |
| Radeon RX 7800 XT | 16 | — | 13 | 35 |
| Radeon RX 9070 XT | 16 | — | 13 | 35.2 |
| Tesla V100 16GB | 16 | — | 13 | 40.5 |
Run Gemma 4 26B A4B with Ollama, llama.cpp or vLLM
The Ollama library has no official tag for it yet.
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve google/gemma-4-26B-A4B-it
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Gemma 4 26B A4B need?
At Q4_K_M with an 8K context Gemma 4 26B A4B needs about 18GB of VRAM; 24GB leaves comfortable headroom.
What is the cheapest GPU that runs Gemma 4 26B A4B?
The Radeon RX 7900 XT: 20GB of VRAM at roughly 83 tokens/s. We have no used-price data for it yet.
Can you run Gemma 4 26B A4B with Ollama?
Not yet: the Ollama library has no official tag for it. You can load the GGUF with llama.cpp instead.