The best local LLM for 8 GB of VRAM

8 GB is the most common amount of graphics memory, and it's enough for a genuinely useful assistant. The trick is picking a model that fits with room left for the conversation.

The pick: Qwen3 8B

At 4-bit it's about 5.2 GB, leaving roughly 2 GB for context on an 8 GB card. It's good at:

• everyday writing, rewriting and summaries,
• light coding questions and explaining code,
• step-by-step answers with its thinking mode (add /think to a message, or /no_think to skip it).

On an RTX 4060 we estimate about 29 tok/s, faster than you can read. On a 3070 or 3060 Ti, closer to 45.

Runner-up: Qwen3 4B for speed

Half the size and twice the speed, with plenty of room for long documents. It's a good pick for scripts, classifying text, or pulling data out of documents, where you run it thousands of times.

What doesn't fit

Gemma 3 12B and Qwen3 14B need 9 to 11 GB. On an 8 GB card they spill onto the CPU and drop to a few tokens per second. You can force them at 3-bit, but Qwen3 8B at 4-bit usually gives better answers than a 14B squeezed that small.

Your card, by the numbers

Here's the RTX 4060 from our calculator:

Other popular 8 GB cards: RTX 3070, RTX 3060 Ti, RTX 4060 Ti 8 GB, RX 7600, Arc A750.

Tips for 8 GB

• Close games and the browser's GPU-heavy tabs. They hold VRAM, and the model will spill onto the CPU.
• Watch your context. Pasting a long document adds memory. If speed suddenly drops, run ollama ps and check the model is still 100% GPU.
• Laptops: a laptop RTX 4060 has less bandwidth than the desktop card, so expect it to be slower.

When you outgrow 8 GB

The next real step up is 12 GB for Qwen3 14B, or 24 GB for the 30B class. The cheapest way there is usually a used card:

• A used RTX 3060 12 GB is around $200.
• A used RTX 3090 is around $700 and runs much bigger models.

See the builds for complete parts lists.