GPUs & Macs › NVIDIA RTX 5070
NVIDIA RTX 5070: local LLM speed
Llama 2 7B (Q4_0) generates 128 tokens/s on the NVIDIA RTX 5070 (CUDA), and processes prompts at 5,783 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 128 | 5,783 | Q4_0 | CUDA FA on | tg128 | llama.cpp #15013 details |
| Llama 2 7B | 128 | 5,185 | Q4_0 | CUDA FA off | tg128 | llama.cpp #15013 details |
| Llama 3.2 1B Instruct | 101 | 8,328 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Llama 3.1 8B Instruct | 55.9 | 2,927 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Qwen2.5 14B Instruct | 33.2 | 1,511 | Q4_K_M | llamafile | LocalScore avg · n=2 | LocalScore LocalScore details |
What fits on the NVIDIA RTX 5070
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 128 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-IQ2_XXS | 7.3 GB | 67.5 estimate |
| gemma-4-12B-it | Q6_K | 9.8 GB | 50.1 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 51.4 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 109 estimate |
| gemma-4-26B-A4B-it | UD-IQ2_M | 10.0 GB | 159 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 51.3 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 59.8 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 114 estimate |
| gemma-4-31B-it | UD-IQ2_XXS | 8.5 GB | 57.4 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 73.0 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 408 estimate |
| gemma-4-E2B-it | Q8_0 | 5.0 GB | 98.7 estimate |
| DeepSeek-R1-0528-Qwen3-8B | Q8_0 | 8.7 GB | 56.3 estimate |
| Llama-3.2-3B-Instruct | Q8_0 | 3.4 GB | 143 estimate |
Similar hardware
NVIDIA TITAN V · NVIDIA TITAN Xp · NVIDIA RTX 3060 · NVIDIA RTX 4070 Ti · NVIDIA RTX 4070 SUPER · NVIDIA RTX 4070 · NVIDIA RTX 3080 Ti · NVIDIA RTX 2080 Ti