Hardware
PC with RTX 5090 32GB
Highest bandwidth in a single card. 32 GB is plenty for 32B, still too tight for 70B Q4.
Memory
32 GB VRAMBandwidth
1792 GB/swww.nvidia.com/en-eu/geforce/graphics-cards/50-series/Price
$8,926€7,973
15 Oct 2026
price unverifiedNot a checked system price. The GPU line is the €6,273 PriceSquirrel figure from 15 Oct 2026. The rest of the box was not rechecked.
Parts in the catalog
- GeForce RTX 5090 32GB
- Ryzen 7 9700X or similar
- 96 GB DDR5
- 1200W PSU, roomy case
Which models fit
Estimate at 8,192 tokens, or the model's maximum when that is shorter. full means the formula fits in usable memory. offload means GPU plus system RAM. no means it does not fit. This is not a measured file.
| Model | Q4estimate | Q5estimate | Q8estimate | FP16estimate |
|---|---|---|---|---|
| Llama 3.2 3B | full | full | full | full |
| Llama 3.1 8B | full | full | full | full |
| Mistral 7B v0.3 | full | full | full | full |
| Gemma 2 9B | full | full | full | full |
| Qwen2.5 14B | full | full | full | offload |
| Phi-4 14B | full | full | full | offload |
| Gemma 2 27B | full | full | offload | offload |
| Qwen2.5 32B | full | full | offload | offload |
| Llama 3.3 70B | offload | offload | offload | no |
| Qwen2.5 72B | offload | offload | offload | no |
| DeepSeek-R1-Distill 70B | offload | offload | offload | no |