Gemma 4 26B-A4B in Q4_K_M with an 8k context takes 17.4 GB of GPU memory. The cheapest option right now is 2× RTX 3060 12 GB, from $0.168/h.
The weight file and the total GPU memory with the context cache. One GPU, one request.
| Quantisation | File | 8K context | 32K context | 128K context | Fits on |
|---|---|---|---|---|---|
| BF16 | 47.0 | 48.6 | 49.1 | 51.0 | RTX PRO 5000 72 GB |
| Q8_0 | 25.0 | 26.6 | 27.1 | 29.0 | Tesla V100 Super 32 GB |
| Q6_K | 21.6 | 23.2 | 23.7 | 25.5 | RTX 3090 24 GB |
| Q5_K_M | 19.7 | 21.3 | 21.8 | 23.6 | RTX 3090 24 GB |
| Q4_K_M | 15.8 | 17.4 | 17.9 | 19.7 | RTX 4000 Ada 20 GB |
| Q3_K_M | 11.9 | 13.5 | 13.9 | 15.8 | RTX A4000 16 GB |
| UD-Q2_K_XL | 9.8 | 11.4 | 11.9 | 13.8 | RTX 3060 12 GB |
How much memory the cache takes per request in FP16. A Q8 cache is about half. 20 KB per token + 1024 × 200 KB.
GPUs from the catalog the model fits on, by price per hour.
The price is the cheapest GPU of the model × the GPU count. A 2–8 GPU server can cost a little more per GPU.
The “Gemma 4 26B” template starts the model in a ready environment: open the link once the server has loaded.
For an API under load there is the vLLM template with an OpenAI-compatible API.
File sizes checked against Hugging Face on 2026-10-19. 1 GB here is 1024³ bytes, like GPU memory. Hugging Face lists decimal gigabytes, so its numbers are 7% larger.