Best LLMs for 12 GB of VRAM
| Text models: 246 | Visual models: 12 | Sized for RTX 3060 12 GB | Strongest first |
Text models for 12 GB
| Model | Best fit | Context | Speed | Strongest at |
|---|---|---|---|---|
| Qwen3.5 9B | 8-bit | 24K | 27 tok/s | Chat, Coding, Agents, Docs |
| Gemma 4 12B | 6-bit | 112K | 25 tok/s | Chat, Coding, Docs |
| Qwen3.5 4B | 8-bit | 176K | 48 tok/s | Agents, Docs |
| Ministral 3 14B | 4-bit | 16K | 30 tok/s | Coding, Agents, Docs |
| Gemma 4 E4B | 8-bit | 128K | 43 tok/s | Chat, Coding, Docs |
| gpt-oss 20B | 3-bit | 40K | 119 tok/s | Agents |
| Ministral 3 8B | 6-bit | 28K | 35 tok/s | Docs |
| Gemma 4 E2B | 8-bit | 128K | 78 tok/s | Chat |
| Nemotron 3 Nano 4B | 8-bit | 256K | 60 tok/s | Docs |
| Llama 3.1 8B | 8-bit | 20K | 31 tok/s | Chat, Coding, Agents, Docs |
| Qwen3 0.6B | 8-bit | 40K | 249 tok/s | Not rated |
| Qwen3 14B | 3-bit | 24K | 35 tok/s | Not rated |
| Qwen3 VL 8B Instruct | 6-bit | 24K | 36 tok/s | Text and images |
| Qwen3 8B | 8-bit | 16K | 30 tok/s | Not rated |
| Qwen2.5 7B Instruct | 8-bit | 32K | 33 tok/s | Not rated |
| Qwen2.5 VL 7B Instruct | 8-bit | 40K | 30 tok/s | Text and images |
| Unlimited OCR | 8-bit | 32K | 181 tok/s | Reading documents |
| Qwen2.5 1.5B Instruct | 8-bit | 32K | 136 tok/s | Not rated |
| Qwen3.5 2B | 8-bit | 256K | 96 tok/s | Text and images |
| Qwen3 VL 4B Instruct | 8-bit | 40K | 51 tok/s | Text and images |
| Qwen3 1.7B | 8-bit | 40K | 106 tok/s | Not rated |
| Qwen3 4B | 8-bit | 40K | 55 tok/s | Not rated |
| Qwen3 4B Instruct 2507 | 8-bit | 44K | 55 tok/s | Not rated |
| Qwen2.5 3B Instruct | 8-bit | 32K | 71 tok/s | Not rated |
| DeepSeek OCR | 8-bit | 8K | 181 tok/s | Reading documents |
| Qwen2.5 Coder 14B Instruct | 3-bit | 20K | 35 tok/s | Coding |
| Qwen2.5 VL 3B Instruct | 8-bit | 125K | 59 tok/s | Text and images |
| Qwen2.5 Coder 7B Instruct | 8-bit | 32K | 33 tok/s | Coding |
| Mistral 7B Instruct V0.2 | 8-bit | 28K | 32 tok/s | Not rated |
| Chandra OCR 2 | 8-bit | 160K | 43 tok/s | Reading documents |
| Qwen3 VL 2B Instruct | 8-bit | 72K | 102 tok/s | Text and images |
| Qwen2 VL 2B Instruct | 8-bit | 32K | 98 tok/s | Text and images |
| MiniCPM5 2B | 8-bit | 128K | 93 tok/s | Not rated |
| Gemma 3 4B IT | 8-bit | 128K | 61 tok/s | Text and images |
| Llama 3.2 3B Instruct | 8-bit | 64K | 70 tok/s | Not rated |
| LightOnOCR 2 1B | 8-bit | 16K | 197 tok/s | Text and images |
| Qwen2.5 14B Instruct | 3-bit | 20K | 35 tok/s | Not rated |
| Qwen3 1.7B Base | 8-bit | 32K | 124 tok/s | Not rated |
| HunyuanOCR | 8-bit | 96K | 184 tok/s | Text and images |
| Qwen3 0.6B Base | 8-bit | 32K | 298 tok/s | Not rated |
| Rax 4.5 | 8-bit | 256K | 96 tok/s | Text and images |
| Qwen2 VL 7B Instruct | 8-bit | 32K | 30 tok/s | Text and images |
| Ornith 1.5 9B | 8-bit | 20K | 27 tok/s | Not rated |
| Unlimited OCR 2 | 8-bit | 120K | 180 tok/s | Reading documents |
| Granite 4.1 3B | 8-bit | 88K | 64 tok/s | Not rated |
| DeepSeek R1 Distill Qwen 1.5B | 8-bit | 128K | 136 tok/s | Not rated |
| Mistral 7B V0.1 | 8-bit | 32K | 32 tok/s | Not rated |
| Qwen2.5 Coder 3B Instruct | 8-bit | 32K | 71 tok/s | Coding |
| Meta Llama 3 8B Instruct Abliterated V3 | 8-bit | 8K | 31 tok/s | Not rated |
| EXAONE 3.5 7.8B Instruct | 8-bit | 20K | 31 tok/s | Not rated |
| Saiga Llama3 8B | 8-bit | 8K | 31 tok/s | Not rated |
| Qwen2.5 Coder 1.5B Instruct | 8-bit | 32K | 136 tok/s | Coding |
| DeepSeek R1 0528 Qwen3 8B | 8-bit | 16K | 30 tok/s | Not rated |
| Gemma 3 12B IT | 5-bit | 28K | 30 tok/s | Text and images |
| Phi 3.5 Mini Instruct | 8-bit | 16K | 60 tok/s | Not rated |
| Qwen3 4B Thinking 2507 | 8-bit | 44K | 55 tok/s | Not rated |
| Phi 4 Mini Instruct | 8-bit | 52K | 58 tok/s | Not rated |
| Dots.ocr | 8-bit | 128K | 79 tok/s | Reading documents |
| DeepSeek R1 Distill Qwen 7B | 8-bit | 52K | 33 tok/s | Not rated |
| Phi 3 Mini 4K Instruct | 8-bit | 4K | 60 tok/s | Not rated |
Fits, context and speed on the RTX 3060 12 GB. A slower card with the same memory holds the same models at lower speed. Ranked by our use-case scores, then by Hugging Face downloads. The top 60 of 246 are shown.
Image and video models
| Model | Best fit | Per output | Strongest at |
|---|---|---|---|
| Stable Diffusion XL 1.0 | 16-bit | 25 s | Images |
| Z-Image Turbo | 8-bit | 36 s | Images |
| FLUX.1 [dev] | 6-bit | 82 s | Images |
| Qwen-Image 2.1 | 8-bit | 9.7 min | Images |
| FLUX.3 Klein | 16-bit | 52 s | Images |
| FLUX.2 [klein] 9B | 8-bit | 17 s | Images |
| Krea 2 Turbo | 5-bit | 81 s | Images |
| Stable Diffusion 3.5 Large | 8-bit | 5.1 min | Images |
| MOSS-Image | 16-bit | 6.3 min | Images |
| Wan 2.2 TI2V-5B | 16-bit | 54 min | Video |
| MOSS Image Base | 16-bit | 6.3 min | Images |
| HunyuanVideo 1.5 | 6-bit | 4.7 h | Video |
Machines with 12 GB
Questions
What is the best LLM for 12 GB of VRAM?
Of the open models we track, the strongest that run well on 12 GB are Qwen3.5 9B, Gemma 4 12B, Qwen3.5 4B. 246 text models fit in all.
Calculated from published specs, not benchmarked. Every machine and tier