Which AI models run on a PC with 16 GB of RAM and an 8 GB graphics card?
20 of Toskar's 21 models run on a PC with 16 GB of RAM and an 8 GB graphics card, and 18 of them run comfortably. A gaming laptop or desktop with an NVIDIA GeForce RTX 4060 or 3070.
Start here
The best model for each job
The largest model of each kind that runs comfortably on this computer.
Everyday chat and writing
Qwen 2.5 14BLargest curated quality pick when your machine has headroom.
Coding
Qwen 2.5 Coder 14BExcellent programming model for larger refactors and multi-file work.
Pictures and screenshots
Qwen 3.5 9BStrong everyday assistant that also sees pictures.
Step-by-step reasoning
DeepSeek R1 Distill 7BStronger reasoning for analysis and multi-step problems.
Every model on this computer
Graded with Toskar 1.12.1's own estimate: the model file plus working memory at an 8K context.
Excellent fitPlenty of memory to spare.
- Qwen 2.5 14B≈ 10.2 GB · Partial GPU offload expected
- Qwen 2.5 Coder 14B≈ 9.7 GB · Partial GPU offload expected
- Qwen 3.5 9B≈ 7.2 GB · Good GPU fit
- Gemma 2 9B≈ 6.5 GB · Good GPU fit
- Qwen 2.5 VL 7B≈ 6.1 GB · Good GPU fit
- Qwen 2.5 Coder 7B≈ 5.3 GB · Good GPU fit
- Qwen 2.5 7B≈ 5.3 GB · Good GPU fit
- DeepSeek R1 Distill 7B≈ 5.3 GB · Good GPU fit
- Mistral 7B Instruct≈ 5.0 GB · Good GPU fit
- Qwen 3.5 4B≈ 3.9 GB · Good GPU fit
- Gemma 3 4B≈ 3.9 GB · Good GPU fit
- Phi 3.5 Mini≈ 3.0 GB · Good GPU fit
- Llama 3.2 3B≈ 2.6 GB · Good GPU fit
- Qwen 2.5 3B≈ 2.5 GB · Good GPU fit
- Gemma 2 2B≈ 2.2 GB · Good GPU fit
- Qwen 2.5 Coder 1.5B≈ 1.6 GB · Good GPU fit
- DeepSeek R1 Distill 1.5B≈ 1.6 GB · Good GPU fit
- Llama 3.2 1B≈ 1.3 GB · Good GPU fit
May run slowlyMay need swap or a shorter context.
- Qwen 3.8 27B≈ 20.4 GB · Partial GPU offload expected
- Muse Glimmer 30B≈ 17.1 GB · Partial GPU offload expected
Too big for this computer: Qwen 2.5 32B. A paired computer with more memory can run them for you.
Questions
- What's the best local AI model for a PC with 16 GB of RAM and an 8 GB graphics card?
- For everyday use, start with Qwen 2.5 14B: it's the largest of its kind that runs comfortably here. For coding, try Qwen 2.5 Coder 14B. For pictures and screenshots, try Qwen 3.5 9B. For step-by-step reasoning, try DeepSeek R1 Distill 7B.
- How big a model can a PC with 16 GB of RAM and an 8 GB graphics card run?
- The largest that runs is Qwen 3.8 27B, at about 20.4 GB; the largest that runs comfortably is Qwen 2.5 14B, at about 10.2 GB. Memory decides it: the model and the conversation have to fit, with room left for everything else you have open.
- Do I need a graphics card?
- It helps a lot. 16 of these models fit entirely in the card's 8.0 GB, which is the fastest way to run them. Larger ones run partly on the card and partly on the processor.
- Does it cost anything to run these models?
- No. Toskar Core is free and open source, and the models download at no cost. Everything runs on your own computer, so there's no subscription and no per-question bill.
Other computers
Mac with Apple silicon
PC with a graphics card
- PC, 16 GB + 8 GB graphics card
- PC, 32 GB + 12 GB graphics card
- PC, 32 GB + 16 GB graphics card
- PC, 64 GB + 24 GB graphics card
Run them with Toskar
Toskar Core is free and open source for Mac, Windows, and Linux. It suggests the models that fit your computer, downloads them for you, and keeps every question on your computer.