Quick Answer: RTX 4070 Ti Super offers 16GB VRAM and starts around current market pricing. It delivers approximately 148 tokens/sec on WeiboAI/VibeThinker-1.5B. It typically draws 285W under load.
This GPU offers reliable throughput for local AI workloads. Pair it with the right model quantization to hit your desired tokens/sec, and monitor prices below to catch the best deal.
Buy directly on Amazon with fast shipping and reliable customer service.
Rotate out primary variants whenever validation flags an issue.
💡 Not ready to buy? Try cloud GPUs first
Test RTX 4070 Ti Super performance in the cloud before investing in hardware. Pay by the hour with no commitment.
| Model | Quantization | Tokens/sec | VRAM used |
|---|---|---|---|
| WeiboAI/VibeThinker-1.5B | Q4 | 147.97 tok/sEstimated Auto-generated benchmark | 1GB |
| deepseek-ai/DeepSeek-OCR | Q4 | 147.97 tok/sEstimated Auto-generated benchmark | 2GB |
| tencent/HunyuanOCR | Q4 | 147.08 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-3-1b-it | Q4 | 146.33 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B | Q4 | 145.24 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-3B-Instruct | Q4 | 145.15 tok/sEstimated Auto-generated benchmark | 2GB |
| allenai/OLMo-2-0425-1B | Q4 | 145.02 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-research/PowerMoE-3b | Q4 | 144.73 tok/sEstimated Auto-generated benchmark | 2GB |
| google/embeddinggemma-300m | Q4 | 143.14 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-2-2b-it | Q4 | 142.97 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-2b | Q4 | 142.79 tok/sEstimated Auto-generated benchmark | 1GB |
| google-bert/bert-base-uncased | Q4 | 142.74 tok/sEstimated Auto-generated benchmark | 1GB |
| nari-labs/Dia2-2B | Q4 | 141.94 tok/sEstimated Auto-generated benchmark | 2GB |
| facebook/sam3 | Q4 | 141.27 tok/sEstimated Auto-generated benchmark | 1GB |
| bigcode/starcoder2-3b | Q4 | 140.18 tok/sEstimated Auto-generated benchmark | 2GB |
| apple/OpenELM-1_1B-Instruct | Q4 | 140.02 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B-Instruct | Q4 | 138.68 tok/sEstimated Auto-generated benchmark | 2GB |
| inference-net/Schematron-3B | Q4 | 136.38 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-1B | Q4 | 136.14 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-3B | Q4 | 133.40 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/Llama-3.2-1B-Instruct | Q4 | 130.31 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-Guard-3-1B | Q4 | 129.07 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-granite/granite-3.3-2b-instruct | Q4 | 127.37 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/Llama-3.2-3B-Instruct | Q4 | 125.96 tok/sEstimated Auto-generated benchmark | 2GB |
| deepseek-ai/deepseek-coder-1.3b-instruct | Q4 | 125.61 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/gemma-3-1b-it | Q4 | 124.93 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | 124.52 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-0.5B | Q4 | 124.47 tok/sEstimated Auto-generated benchmark | 3GB |
| Qwen/Qwen2-1.5B-Instruct | Q4 | 124.18 tok/sEstimated Auto-generated benchmark | 3GB |
| context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16 | Q4 | 123.81 tok/sEstimated Auto-generated benchmark | 2GB |
| Qwen/Qwen2.5-7B-Instruct | Q4 | 123.73 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen3-Embedding-4B | Q4 | 123.67 tok/sEstimated Auto-generated benchmark | 2GB |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-6bit | Q4 | 123.57 tok/sEstimated Auto-generated benchmark | 2GB |
| mistralai/Mistral-7B-Instruct-v0.2 | Q4 | 123.50 tok/sEstimated Auto-generated benchmark | 4GB |
| mistralai/Mistral-7B-Instruct-v0.1 | Q4 | 123.44 tok/sEstimated Auto-generated benchmark | 4GB |
| vikhyatk/moondream2 | Q4 | 123.39 tok/sEstimated Auto-generated benchmark | 4GB |
| unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit | Q4 | 123.15 tok/sEstimated Auto-generated benchmark | 4GB |
| MiniMaxAI/MiniMax-M2 | Q4 | 123.14 tok/sEstimated Auto-generated benchmark | 4GB |
| openai-community/gpt2 | Q4 | 123.13 tok/sEstimated Auto-generated benchmark | 4GB |
| LiquidAI/LFM2-1.2B | Q4 | 123.06 tok/sEstimated Auto-generated benchmark | 1GB |
| black-forest-labs/FLUX.1-dev | Q4 | 122.85 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen3-8B-FP8 | Q4 | 122.82 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2-0.5B | Q4 | 122.64 tok/sEstimated Auto-generated benchmark | 3GB |
| google-t5/t5-3b | Q4 | 122.51 tok/sEstimated Auto-generated benchmark | 2GB |
| ibm-granite/granite-docling-258M | Q4 | 122.47 tok/sEstimated Auto-generated benchmark | 4GB |
| BSC-LT/salamandraTA-7b-instruct | Q4 | 122.45 tok/sEstimated Auto-generated benchmark | 4GB |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | 122.27 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.1-8B-Instruct | Q4 | 122.25 tok/sEstimated Auto-generated benchmark | 4GB |
| EleutherAI/gpt-neo-125m | Q4 | 122.22 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2.5-1.5B-Instruct | Q4 | 122.21 tok/sEstimated Auto-generated benchmark | 3GB |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
| Model | Quantization | Verdict | Estimated speed | VRAM needed |
|---|---|---|---|---|
| meta-llama/Llama-3.2-1B-Instruct | Q8 | Fits comfortably | 104.13 tok/sEstimated | 1GB (have 16GB) |
| meta-llama/Llama-3.2-1B-Instruct | FP16 | Fits comfortably | 54.58 tok/sEstimated | 2GB (have 16GB) |
| openai/gpt-oss-120b | Q4 | Not supported | 24.59 tok/sEstimated | 59GB (have 16GB) |
| openai/gpt-oss-120b | Q8 | Not supported | 15.00 tok/sEstimated | 117GB (have 16GB) |
| openai/gpt-oss-120b | FP16 | Not supported | 8.29 tok/sEstimated | 235GB (have 16GB) |
| Qwen/Qwen2.5-0.5B-Instruct | FP16 | Fits comfortably | 39.07 tok/sEstimated | 11GB (have 16GB) |
| Qwen/Qwen3-32B | Q8 | Not supported | 26.13 tok/sEstimated | 33GB (have 16GB) |
| Qwen/Qwen3-Next-80B-A3B-Instruct | Q8 | Not supported | 17.05 tok/sEstimated | 78GB (have 16GB) |
| Qwen/Qwen3-Next-80B-A3B-Instruct | FP16 | Not supported | 9.00 tok/sEstimated | 156GB (have 16GB) |
| allenai/OLMo-2-0425-1B | Q4 | Fits comfortably | 145.02 tok/sEstimated | 1GB (have 16GB) |
| allenai/OLMo-2-0425-1B | Q8 | Fits comfortably | 89.36 tok/sEstimated | 1GB (have 16GB) |
| google-t5/t5-3b | Q4 | Fits comfortably | 122.51 tok/sEstimated | 2GB (have 16GB) |
| google-t5/t5-3b | Q8 | Fits comfortably | 87.95 tok/sEstimated | 3GB (have 16GB) |
| Qwen/Qwen2.5-14B-Instruct | Q8 | Fits comfortably | 57.66 tok/sEstimated | 14GB (have 16GB) |
| Qwen/Qwen2-0.5B | Q4 | Fits comfortably | 122.64 tok/sEstimated | 3GB (have 16GB) |
| microsoft/phi-4 | Q8 | Fits comfortably | 73.18 tok/sEstimated | 7GB (have 16GB) |
| deepseek-ai/DeepSeek-V3.1 | Q4 | Fits comfortably | 105.75 tok/sEstimated | 4GB (have 16GB) |
| deepseek-ai/DeepSeek-V3.1 | Q8 | Fits comfortably | 73.08 tok/sEstimated | 7GB (have 16GB) |
| deepseek-ai/DeepSeek-V3.1 | FP16 | Fits (tight) | 45.07 tok/sEstimated | 15GB (have 16GB) |
| meta-llama/Llama-3.1-8B | Q4 | Fits comfortably | 117.62 tok/sEstimated | 4GB (have 16GB) |
| meta-llama/Llama-3.1-8B | Q8 | Fits comfortably | 81.64 tok/sEstimated | 9GB (have 16GB) |
| microsoft/DialoGPT-small | FP16 | Fits (tight) | 42.02 tok/sEstimated | 15GB (have 16GB) |
| deepseek-ai/DeepSeek-R1 | Q4 | Fits comfortably | 105.86 tok/sEstimated | 4GB (have 16GB) |
| deepseek-ai/DeepSeek-R1 | Q8 | Fits comfortably | 77.04 tok/sEstimated | 7GB (have 16GB) |
| Qwen/Qwen3-30B-A3B | Q8 | Not supported | 42.74 tok/sEstimated | 31GB (have 16GB) |
| Qwen/Qwen3-30B-A3B | FP16 | Not supported | 23.48 tok/sEstimated | 61GB (have 16GB) |
| IlyaGusev/saiga_llama3_8b | Q4 | Fits comfortably | 106.18 tok/sEstimated | 4GB (have 16GB) |
| IlyaGusev/saiga_llama3_8b | Q8 | Fits comfortably | 80.78 tok/sEstimated | 9GB (have 16GB) |
| Qwen/Qwen3-4B-Base | Q8 | Fits comfortably | 78.67 tok/sEstimated | 4GB (have 16GB) |
| Qwen/Qwen3-4B-Base | FP16 | Fits comfortably | 41.98 tok/sEstimated | 9GB (have 16GB) |
| microsoft/Phi-3.5-mini-instruct | Q4 | Fits comfortably | 119.46 tok/sEstimated | 4GB (have 16GB) |
| Qwen/Qwen2.5-14B | Q8 | Fits comfortably | 61.54 tok/sEstimated | 14GB (have 16GB) |
| Qwen/Qwen2.5-14B | FP16 | Not supported | 34.09 tok/sEstimated | 29GB (have 16GB) |
| google/gemma-2-2b-it | Q4 | Fits comfortably | 142.97 tok/sEstimated | 1GB (have 16GB) |
| google/gemma-2-2b-it | Q8 | Fits comfortably | 95.38 tok/sEstimated | 2GB (have 16GB) |
| google/gemma-2-2b-it | FP16 | Fits comfortably | 55.91 tok/sEstimated | 4GB (have 16GB) |
| Qwen/Qwen2-1.5B-Instruct | Q4 | Fits comfortably | 124.18 tok/sEstimated | 3GB (have 16GB) |
| meta-llama/Llama-Guard-3-1B | Q8 | Fits comfortably | 93.53 tok/sEstimated | 1GB (have 16GB) |
| meta-llama/Llama-Guard-3-1B | FP16 | Fits comfortably | 49.35 tok/sEstimated | 2GB (have 16GB) |
| codellama/CodeLlama-34b-hf | Q4 | Not supported | 36.21 tok/sEstimated | 17GB (have 16GB) |
| codellama/CodeLlama-34b-hf | Q8 | Not supported | 26.66 tok/sEstimated | 35GB (have 16GB) |
| codellama/CodeLlama-34b-hf | FP16 | Not supported | 15.88 tok/sEstimated | 70GB (have 16GB) |
| EleutherAI/gpt-neo-125m | Q4 | Fits comfortably | 122.22 tok/sEstimated | 4GB (have 16GB) |
| EleutherAI/gpt-neo-125m | Q8 | Fits comfortably | 81.06 tok/sEstimated | 7GB (have 16GB) |
| unsloth/Llama-3.2-3B-Instruct | FP16 | Fits comfortably | 48.23 tok/sEstimated | 6GB (have 16GB) |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-8bit | Q8 | Fits comfortably | 80.31 tok/sEstimated | 4GB (have 16GB) |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-8bit | FP16 | Fits comfortably | 42.04 tok/sEstimated | 9GB (have 16GB) |
| ibm-research/PowerMoE-3b | Q4 | Fits comfortably | 144.73 tok/sEstimated | 2GB (have 16GB) |
| ibm-research/PowerMoE-3b | Q8 | Fits comfortably | 103.36 tok/sEstimated | 3GB (have 16GB) |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | Fits comfortably | 124.52 tok/sEstimated | 1GB (have 16GB) |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
Explore how RTX 5070 stacks up for local inference workloads.
Explore how RTX 4060 Ti 16GB stacks up for local inference workloads.
Explore how RX 6800 XT stacks up for local inference workloads.
Explore how RTX 4070 Super stacks up for local inference workloads.
Explore how RTX 3080 stacks up for local inference workloads.