GPUs · NVIDIA
Can the RTX 4070 / 4070 Super run local LLMs?
With 12 GB of VRAM (about 10.8 GB usable) and 504 GB/s of memory bandwidth, the RTX 4070 / 4070 Super runs 16 of 39 popular open models fully on the GPU at 8K context. The largest is Mistral Nemo 12B at Q4_K_M, at roughly 46 tokens per second.
Best picks for the RTX 4070 / 4070 Super
Every model on the RTX 4070 / 4070 Super
Results at 8K context, picking the highest quality quantization that fits. "Offload" means part of the model runs from system RAM (we assume 32 GB), which works but is much slower.
| Model | Status | Quant | Memory | Speed |
|---|---|---|---|---|
| Mistral Nemo 12B | Runs well | Q4_K_M | 9.2 GB | ~46 tok/s |
| Gemma 4 12B | Runs well | Q4_K_M | 11 GB | ~46 tok/s |
| Qwen3.5 9B | Runs well | Q6_K | 9.4 GB | ~43 tok/s |
| Qwen3 8B | Runs well | Q8_0 | 11 GB | ~38 tok/s |
| Llama 3.1 8B | Runs well | Q8_0 | 10 GB | ~39 tok/s |
| DeepSeek-R1 Distill 8B | Runs well | Q8_0 | 10 GB | ~39 tok/s |
| Qwen2.5-Coder 7B | Runs well | Q8_0 | 9.8 GB | ~41 tok/s |
| Mistral 7B v0.3 | Runs well | Q8_0 | 9.4 GB | ~43 tok/s |
| Qwen3.5 4B | Runs well | Q8_0 | 6.4 GB | ~66 tok/s |
| Gemma 3 4B | Runs well | Q8_0 | 6.9 GB | ~72 tok/s |
| Qwen3 4B | Runs well | Q8_0 | 5.8 GB | ~77 tok/s |
| Phi-4 Mini 3.8B | Runs well | Q8_0 | 5.2 GB | ~81 tok/s |
| Llama 3.2 3B | Runs well | Q8_0 | 4.6 GB | ~97 tok/s |
| Llama 3.2 1B | Runs well | Q8_0 | 1.6 GB | ~260 tok/s |
| Gemma 3 1B | Runs well | Q8_0 | 1.4 GB | ~310 tok/s |
| Qwen3 0.6B | Runs well | Q8_0 | 0.9 GB | ~520 tok/s |
| Phi-4 14B | Offload | Q4_K_M | 11 GB | ~5.7 tok/s |
| Gemma 3 12B | Offload | Q4_K_M | 11 GB | ~6.8 tok/s |
| Qwen3 14B | Offload | Q4_K_M | 11 GB | ~5.6 tok/s |
| DeepSeek-R1 Distill 14B | Offload | Q4_K_M | 11 GB | ~5.6 tok/s |
| gpt-oss 20B | Offload | Q4_K_M | 14 GB | ~11 tok/s |
| Codestral 22B | Offload | Q4_K_M | 16 GB | ~3.4 tok/s |
| Devstral Small 24B | Offload | Q4_K_M | 17 GB | ~3.1 tok/s |
| Mistral Small 3 24B | Offload | Q4_K_M | 18 GB | ~3.1 tok/s |
| Gemma 4 26B-A4B (MoE) | Offload | Q4_K_M | 19 GB | ~10 tok/s |
| Qwen3.6 27B | Offload | Q4_K_M | 19 GB | ~2.7 tok/s |
| Qwen3 30B-A3B (MoE) | Offload | Q4_K_M | 21 GB | ~11 tok/s |
| Qwen3-Coder 30B-A3B | Offload | Q4_K_M | 21 GB | ~11 tok/s |
| Gemma 3 27B | Offload | Q4_K_M | 22 GB | ~2.6 tok/s |
| Qwen3 32B | Offload | Q4_K_M | 23 GB | ~2.1 tok/s |
| DeepSeek-R1 Distill 32B | Offload | Q4_K_M | 23 GB | ~2.1 tok/s |
| Qwen2.5-Coder 32B | Offload | Q4_K_M | 23 GB | ~2.1 tok/s |
| Qwen3.6 35B-A3B (MoE) | Offload | Q4_K_M | 23 GB | ~12 tok/s |
| Gemma 4 31B | Offload | Q4_K_M | 25 GB | ~2.3 tok/s |
| Llama 3.3 70B | Won’t fit | — | 48 GB | — |
| DeepSeek-R1 Distill 70B | Won’t fit | — | 48 GB | — |
| Llama 4 Scout 109B-A17B | Won’t fit | — | 73 GB | — |
| gpt-oss 120B | Won’t fit | — | 77 GB | — |
| Qwen3 235B-A22B (MoE) | Won’t fit | — | 155 GB | — |
What this means in practice
The RTX 4070 / 4070 Super is comfortable with models up to about 12B to 14B parameters. gpt-oss 20B is the most useful step up beyond that: it needs about 14 GB, so part of it runs from system RAM, but as a mixture-of-experts model it stays reasonably quick at around 11 tokens per second. Models from Llama 3.3 70B upward need 48 GB or more and are out of reach. NVIDIA cards have the broadest support: Ollama, LM Studio and llama.cpp all use CUDA out of the box.
The next memory tier up is 16 GB: RTX 5080, RTX 5070 Ti, RTX 4080 / 4080 Super. Other 12 GB cards: RTX 5070, RTX 3060, RTX 4080 Laptop, RX 7700 XT, RX 6700 XT, Arc B580.
Compare the RTX 4070 / 4070 Super
Questions
What is the biggest model the RTX 4070 / 4070 Super can run?
At 8K context the largest model that fits entirely in its 12 GB of VRAM is Mistral Nemo 12B at Q4_K_M, at roughly 46 tokens per second.
Can the RTX 4070 / 4070 Super run 70B models like Llama 3.3 70B?
No. It needs about 48 GB at Q4, which is far beyond 12 GB even with system RAM offload.
What is the best model for coding on the RTX 4070 / 4070 Super?
Qwen3.5 9B at Q6_K fits fully and runs at about 43 tokens per second. Pair it with an editor extension like Continue pointed at Ollama.
Estimates based on memory size and bandwidth; real speed varies with drivers, software and context length. See how we calculate VRAM.