Consumer · GDDR6X
RTX 3090 24 GB (used): which LLMs can it run?
On its own, it loads up to DeepSeek R1 Distill 32B in Q4. With several cards, memory adds up and opens bigger models.
Specifications
Memory24 GB
Bandwidth936 GB/s
FP16 compute~71 TF
Power350 W
Indicative price€750
Cloud rental€0.25/h
Against cards at the same price
Models and expected speed
| Model | Q4 | Q8 | FP16 |
|---|---|---|---|
| Qwen3 4B | ✓ ~202 tok/s | ✓ ~117 tok/s | ✓ ~63 tok/s |
| Llama 3.1 8B | ✓ ~104 tok/s | ✓ ~60 tok/s | ✓ ~32 tok/s |
| Qwen3 8B | ✓ ~102 tok/s | ✓ ~58 tok/s | ✓ ~31 tok/s |
| Gemma 3 12B | ✓ ~70 tok/s | ✓ ~40 tok/s | 2 × · ~32 tok/s |
| Mistral Nemo 12B | ✓ ~69 tok/s | ✓ ~39 tok/s | 2 × · ~31 tok/s |
| Phi-4 14B | ✓ ~57 tok/s | ✓ ~33 tok/s | 2 × · ~26 tok/s |
| Qwen3 14B | ✓ ~57 tok/s | ✓ ~33 tok/s | 2 × · ~26 tok/s |
| gpt-oss-20b | ✓ ~223 tok/s | 2 × · ~195 tok/s | 2 × · ~105 tok/s |
| Mistral Small 24B | ✓ ~35 tok/s | 2 × · ~30 tok/s | 3 × · ~24 tok/s |
| Gemma 3 27B | ✓ ~31 tok/s | 2 × · ~27 tok/s | 3 × · ~21 tok/s |
| Qwen3.8 27B | ✓ ~30 tok/s | 2 × · ~26 tok/s | 3 × · ~21 tok/s |
| Qwen3 30B-A3B | ✓ ~242 tok/s | 2 × · ~212 tok/s | 3 × · ~172 tok/s |
| Qwen3 32B | ✓ ~26 tok/s | 2 × · ~22 tok/s | 4 × · ~23 tok/s |
| Qwen2.5 Coder 32B | ✓ ~26 tok/s | 2 × · ~22 tok/s | 4 × · ~23 tok/s |
| DeepSeek R1 Distill 32B | ✓ ~26 tok/s | 2 × · ~22 tok/s | 4 × · ~23 tok/s |
| Mixtral 8x7B | 2 × · ~98 tok/s | 3 × · ~84 tok/s | ✗ |
| Llama 3.3 70B | 3 × · ~27 tok/s | 4 × · ~21 tok/s | ✗ |
| GLM-4.5-Air 106B | 3 × · ~158 tok/s | ✗ | ✗ |
| gpt-oss-120b | 4 × · ~481 tok/s | ✗ | ✗ |
| Mistral Large 123B | 4 × · ~21 tok/s | ✗ | ✗ |
| Qwen3 235B-A22B | ✗ | ✗ | ✗ |
| GLM-4.5 355B | ✗ | ✗ | ✗ |
| Llama 3.1 405B | ✗ | ✗ | ✗ |
| DeepSeek R1 671B | ✗ | ✗ | ✗ |
| DeepSeek V3.1 671B | ✗ | ✗ | ✗ |
| Kimi K2 1T | ✗ | ✗ | ✗ |
8k tokens, one user. “2 ×” = number of cards needed, speed with tensor parallelism (vLLM). ✗ = doesn’t fit, even with 4 cards.