Qwen3 30B-A3B in Q4_K_M with an 8k context takes 19.1 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 | Fits on |
|---|---|---|---|---|
| BF16 | 56.9 | 58.7 | 60.9 | RTX PRO 5000 72 GB |
| FP8 | 30.2 | 32.0 | 34.3 | A100 PCIe 40 GB |
| Q8_0 | 30.3 | 32.0 | 34.3 | A100 PCIe 40 GB |
| Q6_K | 23.4 | 25.2 | 27.4 | Tesla V100 Super 32 GB |
| Q5_K_M | 20.2 | 22.0 | 24.3 | RTX 3090 24 GB |
| Q4_K_M | 17.3 | 19.1 | 21.3 | RTX 4000 Ada 20 GB |
| Q3_K_M | 13.7 | 15.5 | 17.7 | RTX A4000 16 GB |
| Q2_K | 10.5 | 12.3 | 14.5 | RTX A4000 16 GB |
How much memory the cache takes per request in FP16. A Q8 cache is about half. 96 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 qwen3:30bFor 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.