Which AI models can the NVIDIA RTX 5090 run?
- 32 GB VRAM
- 1792 GB/s bandwidth
- NVIDIA
Every model we track, checked against the NVIDIA RTX 5090 at 8K context.
| Model | Runs? | Best fit |
|---|---|---|
| Llama 3.2 1B | Yes | FP16 · ~3.3 GB · ~434 tok/s |
| Llama 3.2 3B | Yes | FP16 · ~8 GB · ~167 tok/s |
| Llama 3.1 8B | Yes | FP16 · ~17.7 GB · ~67 tok/s |
| Llama 3.1 70B | Yes, with heavy quantization | Q2_K · ~29.8 GB · ~41 tok/s |
| Qwen2.5 7B | Yes | FP16 · ~16.3 GB · ~71 tok/s |
| Qwen2.5 14B | Yes | FP16 · ~31.6 GB · ~37 tok/s |
| Qwen2.5 32B | Yes | Q6_K · ~29.6 GB · ~40 tok/s |
| Qwen2.5 72B | Yes, with heavy quantization | Q2_K · ~30.5 GB · ~39 tok/s |
| Qwen3 8B | Yes | FP16 · ~18.2 GB · ~66 tok/s |
| Qwen3 30B-A3B | Yes | Q6_K · ~26.4 GB · ~397 tok/s |
| DeepSeek R1 Distill Qwen 32B | Yes | Q6_K · ~29.6 GB · ~40 tok/s |
| Mistral 7B v0.3 | Yes | FP16 · ~16.2 GB · ~74 tok/s |
| Mistral Nemo 12B | Yes | FP16 · ~26.3 GB · ~44 tok/s |
| Mistral Small 24B | Yes | Q8_0 · ~27 GB · ~43 tok/s |
| Mixtral 8x7B | Yes | Q4_K_M · ~30 GB · ~137 tok/s |
| Gemma 2 9B | Yes | FP16 · ~21.9 GB · ~58 tok/s |
| Gemma 2 27B | Yes | Q6_K · ~26 GB · ~48 tok/s |
| Gemma 3 4B | Yes | FP16 · ~10.3 GB · ~125 tok/s |
| Gemma 3 12B | Yes | FP16 · ~28.2 GB · ~44 tok/s |
| Gemma 3 27B | Yes | Q6_K · ~27.2 GB · ~48 tok/s |
| Phi-3.5 Mini 3.8B | Yes | FP16 · ~11.5 GB · ~141 tok/s |
| Phi-4 14B | Yes | FP16 · ~31.7 GB · ~37 tok/s |
| Qwen3 4B | Yes | FP16 · ~9.8 GB · ~134 tok/s |
| Qwen3 14B | Yes | FP16 · ~31.5 GB · ~36 tok/s |
| Qwen3 32B | Yes | Q6_K · ~29.6 GB · ~40 tok/s |
| Qwen3 Coder 30B-A3B | Yes | Q6_K · ~26.4 GB · ~397 tok/s |
| gpt-oss 20B | Yes | Q8_0 · ~23.2 GB · ~281 tok/s |
| gpt-oss 120B | No | Needs ~72 GB at Q4_K_M; 32 GB available |
| Llama 4 Scout 17B-16E | No | Needs ~68.3 GB at Q4_K_M; 32 GB available |
Run it in the cloud
Need more than 32 GB? Rent a bigger GPU by the hour.
Rent a cloud GPU by the hour. No upfront cost, cancel any time.