GPUs & Macs › NVIDIA RTX 3060 Ti
NVIDIA RTX 3060 Ti: local LLM speed
Llama 2 7B (Q4_0) generates 90.3 tokens/s on the NVIDIA RTX 3060 Ti (Vulkan), and processes prompts at 2,645 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 90.3 | 2,645 | Q4_0 | Vulkan FA on | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 86.7 | 2,361 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
| Llama 3.2 1B Instruct | 193 | 7,644 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Llama 3.1 8B Instruct | 60.2 | 1,839 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Qwen2.5 14B Instruct | 7.5 | 430 | Q4_K_M | llamafile | LocalScore avg · part CPU | LocalScore details |
What fits on the NVIDIA RTX 3060 Ti
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 90.3 tok/s (Vulkan, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| gemma-4-12B-it | IQ4_XS | 6.4 GB | 54.2 estimate |
| Qwen3.5-9B | Q5_K_S | 6.4 GB | 54.3 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 77.1 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q4_K_M | 5.8 GB | 59.1 estimate |
| gemma-4-E4B-it | Q5_K_M | 5.5 GB | 63.0 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 80.7 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 288 estimate |
| gemma-4-E2B-it | Q8_0 | 5.0 GB | 69.5 estimate |
| DeepSeek-R1-0528-Qwen3-8B | Q5_K_M | 5.9 GB | 59.0 estimate |
| Llama-3.2-3B-Instruct | Q8_0 | 3.4 GB | 101 estimate |
| Phi-4-mini-instruct | Q8_0 | 4.1 GB | 84.6 estimate |
| gemma-3-4b-it | Q8_0 | 4.1 GB | 83.6 estimate |
| Qwen3-14B | UD-IQ3_XXS | 6.0 GB | 57.4 estimate |
| gemma-3-12b-it | Q3_K_M | 6.0 GB | 57.5 estimate |
Similar hardware
NVIDIA RTX 2070 SUPER · NVIDIA RTX 4060 Ti · NVIDIA RTX 2060 SUPER · NVIDIA GTX 1070 Ti · NVIDIA RTX 3070 · NVIDIA GTX 1070 · NVIDIA RTX 3070 Ti · NVIDIA RTX 2080 SUPER