Qwen3.8 27B in Q4_K_M with an 8k context takes 17.2 GB of GPU memory. The cheapest option right now is 2× RTX 3060 12 GB, from $0.174/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 | 50.9 | 52.8 | 54.3 | 60.3 | RTX PRO 5000 72 GB |
| FP8 | 28.7 | 30.6 | 32.1 | 38.1 | Tesla V100 32 GB |
| Q8_0 | 27.1 | 28.9 | 30.4 | 36.4 | Tesla V100 32 GB |
| Q6_K | 20.5 | 22.3 | 23.8 | 29.8 | RTX 3090 24 GB |
| Q5_K_M | 18.4 | 20.3 | 21.8 | 27.8 | RTX 3090 24 GB |
| Q4_K_M | 15.3 | 17.2 | 18.7 | 24.7 | RTX 4000 Ada 20 GB |
| UD-Q2_K_XL | 9.2 | 11.0 | 12.5 | 18.5 | RTX 3060 12 GB |
How much memory the cache takes per request in FP16. A Q8 cache is about half. 64 KB per token.
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 “Qwen 3.8” template starts the model in a ready environment: open the link once the server has loaded.
In the Ollama template the model is one command away:
ollama run qwen3.8:27bFor 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.