What can the NVIDIA GeForce RTX 5070 Ti run?
16 GB VRAM (~15.2 GB usable for models) · 896 GB/s memory bandwidth · launch MSRP $749
Measured: 176.85 tok/s on Llama 2 7B Q4_0 (llama.cpp tg128). Source: llama.cpp CUDA scoreboard (Llama 2 7B Q4_0, tg128, no FA)
| Model | Verdict | Best quant that fits | Est. tok/s (Q4) |
|---|---|---|---|
| Llama 3.1 8B Instruct | Yes | Q8_0 | ~139 |
| Llama 3.3 70B Instruct | No | — | — |
| Llama 3.2 3B Instruct | Yes | F16 | ~300 |
| Llama 4 Scout (17B active, 16 experts) | No | — | — |
| Qwen3 8B | Yes | Q8_0 | ~136 |
| Qwen3 14B | Yes | Q6_K | ~75.3 |
| Qwen3 32B | No | — | — |
| Qwen3 30B-A3B (MoE) | No | — | — |
| Qwen3 235B-A22B (MoE) | No | — | — |
| Qwen2.5 Coder 32B Instruct | No | — | — |
| Gemma 3 4B | Yes | F16 | ~259 |
| Gemma 3 12B | Yes | Q6_K | ~91.4 |
| Gemma 3 27B | No | — | — |
| Mistral Small 3.1 24B | Yes | Q3_K_M | ~57.8 |
| Mistral Nemo 12B | Yes | Q8_0 (4K ctx) | ~91.4 |
| Phi-4 14B | Yes | Q6_K | ~75.9 |
| DeepSeek R1 Distill Qwen 32B | No | — | — |
| DeepSeek R1 Distill Llama 70B | No | — | — |
| gpt-oss-20b (MoE) | Yes | MXFP4 (native) | ~163 |
| gpt-oss-120b (MoE) | No | — | — |
| DeepSeek V3 / R1 671B (MoE) | No | — | — |
Estimated tok/s = calibrated bandwidth ÷ active weight size. Real results vary with runtime, context and drivers.
No suitable GPU? Rent one by the hour on Vast.ai or RunPod (affiliate links).