GPUs & Macs › NVIDIA RTX 3090 Ti
NVIDIA RTX 3090 Ti: local LLM speed
Llama 2 7B (Q4_0) generates 172 tokens/s on the NVIDIA RTX 3090 Ti (CUDA), and processes prompts at 6,924 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 172 | 6,924 | Q4_0 | CUDA FA on | tg128 | llama.cpp #15013 details |
| Llama 2 7B | 171 | 6,567 | Q4_0 | CUDA FA off | tg128 | llama.cpp #15013 details |
| Llama 3.2 1B Instruct | 354 | 15,120 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Llama 3.1 8B Instruct | 110 | 4,024 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Qwen2.5 14B Instruct | 64.2 | 2,202 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
What fits on the NVIDIA RTX 3090 Ti
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 172 tok/s (CUDA, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | UD-Q6_K | 22.0 GB | 30.0 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 52.0 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 69.1 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 147 estimate |
| Qwen3.6-35B-A3B | UD-Q4_K_S | 20.9 GB | 184 estimate |
| gemma-4-26B-A4B-it | UD-Q5_K_M | 21.2 GB | 101 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 69.0 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 80.4 estimate |
| Muse-Glimmer-30B | UD-Q5_K_M | 19.2 GB | 34.3 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 154 estimate |
| gemma-4-31B-it | Q5_K_M | 21.7 GB | 30.4 estimate |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 14.2 GB | 150 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 98.0 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 549 estimate |
Similar hardware
NVIDIA RTX 4090 D · NVIDIA RTX 4090 · NVIDIA RTX 3090 · NVIDIA TITAN RTX · NVIDIA Quadro P6000 · NVIDIA RTX 5090 · NVIDIA RTX 5080 · NVIDIA RTX 5070 Ti