The best local AI models for an 8 GB graphics card

8 GB is the most common amount of VRAM in gaming PCs and laptops: RTX 3050, 3060 Ti, 3070, 4060, laptop 4060 and 4070 all ship with it. It is also enough to run useful models, as long as you choose well. Here is what we would install, by job.

Everyday chat and writing: Qwen3 8B

At Q4 it takes about 5.5 GB with 8K context and runs at roughly 35 to 50 tokens per second on a desktop 8 GB card. It follows instructions well, writes clean English and handles summarising, rewriting and simple questions. Llama 3.1 8B is the alternative if you prefer its more conversational tone.

Coding: Qwen2.5-Coder 7B

Purpose-trained on code, it is noticeably better at completions and small refactors than general 8B models. Pair it with the Continue extension in VS Code for local autocomplete. If you can stretch context to 16K, keep quantization at Q4 to leave room.

Reasoning: DeepSeek-R1 Distill 8B

This one thinks out loud before answering, which makes it slower per answer but far better at maths and multi-step logic than plain chat models of the same size. Keep context at 8K or more because its reasoning traces are long.

Images: Gemma 3 4B

The smallest model here that can look at a picture and describe it. At Q4 it is under 4 GB, leaving plenty of space for the vision encoder and a long context.

What not to bother with

14B models at Q4 technically load on 8 GB but leave almost no room for context, and 24B to 32B models will spill into system RAM and run at a few tokens per second. If you really want a bigger model, Qwen3 30B-A3B is the one exception: it is a mixture-of-experts model that only activates 3B parameters per token, so it runs fast even when partly offloaded. It needs about 18 GB total, so it will use your system RAM, but speed stays usable.

Check the exact numbers for your card in the checker, especially if you have a laptop variant with lower bandwidth.