Can the RTX 5090 run Gemma 4 26B A4B?
Yes. Gemma 4 26B A4B fits the RTX 5090's 32 GB at Q4_K_M (about 16.8 GB needed), running at an estimated 303 tok/s. There's room for up to Q8_0 if you want higher quality.
Every quantization
Gemma 4 26B A4B (26.5B, 4B active) on 32 GB
| Quant | Weights | Needs | Fit | Est. speed |
|---|---|---|---|---|
| Q4_K_M | 15 GB | 16.8 GB | Fits easily | ~303 tok/s |
| Q5_K_M | 19 GB | 20.8 GB | Fits | ~239 tok/s |
| Q8_0 | 28 GB | 29.8 GB | Tight | ~162 tok/s |
| FP16 | 53 GB | 54.8 GB | Too big | - |
Estimates. Needs = weights + 1.8 GB for KV cache and runtime at modest context. Speed is single-stream decode scaled from the RTX 5090's 1792 GB/s bandwidth (Estimated from memory bandwidth (llama.cpp decode)). Long contexts need more memory and run slower.
Run it
Copy, paste, go
llama-server -hf google/gemma-4-26B-A4B-it-qat-q4_0-gguf -ngl 99Ollama pulls a Q4 build by default. Pick a larger quant only if the table above says it fits. Full Gemma 4 26B A4B details →
Compare
Same model, other hardware - and other models on this card
Gemma 4 26B A4B on other hardware
- RTX 5090 D~302 tok/s
- RTX 6000D~265 tok/s
- RTX 4090~184 tok/s
- Radeon RX 7900 XTX~162 tok/s
- RTX 3090~161 tok/s
- MacBook Pro 16" M4 Max~92 tok/s
- Arc Pro B60~77 tok/s
- DGX Spark (GB10)~46 tok/s
Also runs on the RTX 5090
- Qwen3.6 35B A3B~391 tok/s
- Qwen3.5 35B A3B~391 tok/s
- dolphin-2.9.1-yi-1.5-34b~51 tok/s
- Gemma 4 31B~54 tok/s
- Qwen3 32B~54 tok/s
- GLM-4.7-Flash~396 tok/s
- Qwen3-30B-A3B-Instruct-2507~387 tok/s
- Qwen3.6 27B~64 tok/s
Frequently asked
Quick answers
Can the NVIDIA RTX 5090 run Gemma 4 26B A4B?
Yes. Gemma 4 26B A4B fits the RTX 5090's 32 GB at Q4_K_M (about 16.8 GB needed), running at an estimated 303 tok/s. There's room for up to Q8_0 if you want higher quality.
How much VRAM does Gemma 4 26B A4B need?
About 15 GB of weights at Q4_K_M, 28 GB at Q8_0 and 53 GB at FP16. Add roughly 1.8 GB for the KV cache and runtime at modest context lengths.
How fast is Gemma 4 26B A4B on the RTX 5090?
Roughly 303 tokens per second at Q4_K_M (single user, llama.cpp-style decode). This is an estimate from the RTX 5090's 1792 GB/s memory bandwidth; as a mixture-of-experts model only about 4B parameters are read per token, which is why it is faster than its size suggests.