Qwen3 32B in Q4_K_M with an 8k context takes 21.4 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 | Fits on |
|---|---|---|---|---|
| BF16 | 61.0 | 64.1 | 70.1 | RTX PRO 5000 72 GB |
| FP8 | 32.0 | 35.0 | 41.0 | A100 PCIe 40 GB |
| Q8_0 | 32.4 | 35.5 | 41.5 | A100 PCIe 40 GB |
| Q6_K | 25.0 | 28.1 | 34.1 | Tesla V100 32 GB |
| Q5_K_M | 21.6 | 24.7 | 30.7 | Tesla V100 32 GB |
| Q4_K_M | 18.4 | 21.4 | 27.4 | RTX 3090 24 GB |
| Q3_K_M | 14.9 | 17.9 | 23.9 | RTX 4000 Ada 20 GB |
| Q2_K | 11.5 | 14.5 | 20.5 | RTX A4000 16 GB |
How much memory the cache takes per request in FP16. A Q8 cache is about half. 256 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.
In the Ollama template the model is one command away:
ollama run qwen3:32bFor 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.