AI Hashrate · RTX 5090 32GB
Which LLM should you run on the RTX 5090 32GB?
32 GB · 1792 GB/s · 575 W
Every number below comes from the same engine behind the wizard: VRAM fit per quantisation, decode speed from memory bandwidth, and a command you can paste. Nothing is scraped or invented.
The short answer
gemma-4-31B-it — about 54.8 tok/s decode
Also fits
| Model | Q4 VRAM | Est. tok/s |
|---|---|---|
| gemma-4-26B-A4B-it | 15.4 GB | 443.1 |
| Qwen3.6-35B-A3B | 21.3 GB | 561.3 |
| Dolphin3.0-R1-Mistral-24B | 15.7 GB | 70.2 |
| Qwen3.8-27B-Uncensored | 16.9 GB | 61.2 |
| Qwen3.5-35B-A3B | 20.6 GB | 561.3 |
| Qwen2.5-32B-Instruct-abliterated | 21.4 GB | 51.3 |
| Cydonia-24B-v4.1 | 15.7 GB | 70.2 |
Run it
llama-server \
-hf <GGUF_REPO>:Q4_K_M \
-c 8192 \
-ngl 99 \
--host 127.0.0.1 --port 8080These figures are modelled estimates from hardware specs and public reports, not benchmarks measured on this card. The wizard shows where every number comes from.