Can I Run It? Which LLMs Your GPU Can Run
Choose your graphics card or Mac to see every model that fits, how fast it runs and the highest-quality quantization that fits. An RTX 3060 12GB runs 25 of 58 models fully at Q4_K_M with 8K context.
25 of 58 models fit
RTX 3060 12GB (12.0 GB) at Q4_K_M with 8K context. Largest that fits fully: Qwen 3 14B.
| Model | Needs | Status | Speed | Best quant that fits |
|---|---|---|---|---|
| Qwen 2.5 14B | 10.7 GB | Fits | ~24 tok/s | Q4_K_M |
| DeepSeek R1 Distill Qwen 14B | 10.7 GB | Fits | ~24 tok/s | Q4_K_M |
| Phi-4 14B | 10.6 GB | Fits | ~24 tok/s | Q4_K_M |
| Qwen 3 14B | 10.5 GB | Fits | ~24 tok/s | Q4_K_M |
| Ministral 3 14B | 9.7 GB | Fits | ~26 tok/s | Q5_K_M |
| Mistral Nemo 12B | 9.0 GB | Fits | ~29 tok/s | Q5_K_M |
| Gemma 3 12B | 8.8 GB | Fits | ~29 tok/s | Q6_K |
| Ministral 3 8B | 6.7 GB | Fits | ~40 tok/s | Q8_0 |
| Qwen 3 8B | 6.6 GB | Fits | ~41 tok/s | Q8_0 |
| Qwen 3.5 9B | 6.4 GB | Fits | ~41 tok/s | Q8_0 |
| Llama 3.1 8B | 6.3 GB | Fits | ~43 tok/s | Q8_0 |
| DeepSeek R1 Distill Llama 8B | 6.3 GB | Fits | ~43 tok/s | Q8_0 |
| Mistral 7B v0.3 | 5.7 GB | Fits | ~47 tok/s | Q8_0 |
| Gemma 4 E4B | 5.7 GB | Fits | ~46 tok/s | Q8_0 |
| Qwen 2.5 7B | 5.6 GB | Fits | ~48 tok/s | Q8_0 |
| DeepSeek R1 Distill Qwen 7B | 5.6 GB | Fits | ~48 tok/s | Q8_0 |
| Qwen 3 4B | 4.2 GB | Fits | ~75 tok/s | FP16 / BF16 |
| Phi-4 Mini | 4.1 GB | Fits | ~80 tok/s | FP16 / BF16 |
| Gemma 4 E2B | 4.0 GB | Fits | ~73 tok/s | FP16 / BF16 |
| Gemma 3 4B | 3.7 GB | Fits | ~81 tok/s | FP16 / BF16 |
| Qwen 3.5 4B | 3.7 GB | Fits | ~83 tok/s | FP16 / BF16 |
| Ministral 3 3B | 3.5 GB | Fits | ~91 tok/s | FP16 / BF16 |
| Llama 3.2 3B | 3.5 GB | Fits | ~94 tok/s | FP16 / BF16 |
| Llama 3.2 1B | 1.7 GB | Fits | ~253 tok/s | FP16 / BF16 |
| Gemma 3 1B | 1.6 GB | Fits | ~336 tok/s | FP16 / BF16 |
| Mixtral 8x7B | 28.0 GB | 40% on GPU | ~7.3 tok/s | – |
| Qwen 2.5 32B | 21.4 GB | 53% on GPU | ~3.5 tok/s | – |
| Qwen 2.5 Coder 32B | 21.4 GB | 53% on GPU | ~3.5 tok/s | – |
| DeepSeek R1 Distill Qwen 32B | 21.4 GB | 53% on GPU | ~3.5 tok/s | – |
| Qwen 3 32B | 21.4 GB | 53% on GPU | ~3.5 tok/s | – |
| Qwen 3.5 35B-A3B | 21.0 GB | 54% on GPU | ~33 tok/s | – |
| Qwen 3.6 35B-A3B | 21.0 GB | 54% on GPU | ~33 tok/s | – |
| Gemma 4 31B | 19.8 GB | 57% on GPU | ~4.0 tok/s | – |
| Qwen 3 30B-A3B | 18.8 GB | 60% on GPU | ~30 tok/s | – |
| Qwen3-Coder 30B-A3B | 18.8 GB | 60% on GPU | ~30 tok/s | – |
| Gemma 3 27B | 17.6 GB | 64% on GPU | ~5.1 tok/s | – |
| Qwen 3.5 27B | 16.9 GB | 67% on GPU | ~5.5 tok/s | Q2_K |
| Qwen 3.6 27B | 16.9 GB | 67% on GPU | ~5.5 tok/s | Q2_K |
| Qwen 3.8 27B | 16.9 GB | 67% on GPU | ~5.5 tok/s | Q2_K |
| Mistral Small 3.1 24B | 15.6 GB | 72% on GPU | ~6.7 tok/s | Q2_K |
| Gemma 4 26B-A4B | 15.6 GB | 72% on GPU | ~35 tok/s | Q2_K |
| gpt-oss-20b | 12.9 GB | 87% on GPU | ~56 tok/s | Q3_K_M |
| Kimi K2.6 | 566 GB | Too small | – | – |
| DeepSeek R1 671B | 380 GB | Too small | – | – |
| Llama 3.1 405B | 234 GB | Too small | – | – |
| Qwen 3.5 397B-A17B | 226 GB | Too small | – | – |
| DeepSeek V4 Flash | 161 GB | Too small | – | – |
| Qwen 3 235B-A22B | 135 GB | Too small | – | – |
| MiniMax-M2 | 133 GB | Too small | – | – |
| Qwen 3.5 122B-A10B | 70.2 GB | Too small | – | – |
| Mistral Small 4 119B | 68.2 GB | Too small | – | – |
| gpt-oss-120b | 67.2 GB | Too small | – | – |
| Llama 4 Scout | 64.0 GB | Too small | – | – |
| GLM-4.5-Air | 62.1 GB | Too small | – | – |
| Qwen3-Next 80B-A3B | 46.3 GB | Too small | – | – |
| Qwen 2.5 72B | 44.5 GB | Too small | – | – |
| Llama 3.3 70B | 43.2 GB | Too small | – | – |
| DeepSeek R1 Distill Llama 70B | 43.2 GB | Too small | – | – |
How to read the results
- Fits: the whole model and its context fit in VRAM with a safety margin.
- Tight fit: it fits, but only if nothing else uses the GPU. Close the browser or drive your display from integrated graphics.
- % on GPU: the rest of the model runs from system RAM (llama.cpp, Ollama and LM Studio do this automatically). Expect a large slowdown.
- Best quant that fits: the highest-quality GGUF format that fits fully at your chosen context.
Don't see your GPU?
Choose Custom VRAM and enter your memory size and bandwidth (from the spec sheet). Bandwidth sets the speed estimate; memory sets what fits.
Want the reverse view? The LLM VRAM calculator starts from a model and lists GPUs, and the VRAM tier pages rank the best models for each memory size.
Frequently asked questions
How accurate is "Can I Run It"?
Memory estimates are usually within about 5–10% of what llama.cpp reports, because they use each model's real layer and head counts. Speed estimates are rougher (±25%) and assume single-user generation.
Can I use system RAM if my VRAM is too small?
Yes. Runtimes like llama.cpp and Ollama keep some layers on the CPU. It works, but every token then waits on system RAM, which is 10–30× slower than GPU memory, so speed drops sharply once more than a few layers are offloaded.
What about AMD and Intel GPUs?
They run llama.cpp-based apps through ROCm, Vulkan or SYCL. Memory needs are the same; the speed estimate assumes slightly lower efficiency than NVIDIA.