- Learn
- Pick a model
The best local LLM for 12 GB of VRAM (RTX 3060, 4070, B580)
What to run on a 12 GB graphics card like the RTX 3060, 4070 or Arc B580. Picks for smarts, images and speed, with estimated speeds for each card.
12 GB is the first size where a local model starts to feel smart rather than just handy. It fits the 12B to 14B class, which is a clear step up from the 8B models that 8 GB cards are stuck with.
The pick: Qwen3 14B
ollama run qwen3:14bAt 4-bit it's about 9.3 GB, which leaves roughly 1.5 GB for the conversation on a 12 GB card. It's noticeably better than 8B models at:
- coding help and reviewing code,
- reasoning through a problem with thinking mode (add
/thinkto a message,/no_thinkto skip it), - following longer, more detailed instructions,
- tool calling and simple agents.
On a used RTX 3060 12 GB we estimate about 21 tok/s, a bit faster than most people read. On an RTX 4070, about 30.
For images: Gemma 3 12B
ollama run gemma3:12bGoogle's model reads images as well as text: screenshots, charts, photos of receipts. It's also strong at writing and at languages other than English, and it has a long 128K context window. At about 8.1 GB it leaves more room for context than Qwen3 14B. It's weaker at hard reasoning and math, where Qwen3 14B wins. See the best local vision model.
For speed: Qwen3 8B
ollama run qwen3:8bIf you want snappy answers or long documents, Qwen3 8B runs about twice as fast as the 14B and leaves around 5 GB for context. On a 3060 that's an estimated 38 tok/s.
Card by card
Speed depends on memory bandwidth (see why bandwidth sets speed), so 12 GB cards vary a lot:
| Card | Bandwidth | Qwen3 14B | Gemma 3 12B | Qwen3 8B |
|---|---|---|---|---|
| RTX 3060 12 GB | 360 GB/s | ~21 tok/s | ~24 tok/s | ~38 tok/s |
| RX 6700 XT | 384 GB/s | ~23 tok/s | ~26 tok/s | ~41 tok/s |
| Arc B580 | 456 GB/s | ~27 tok/s | ~31 tok/s | ~48 tok/s |
| RTX 4070 | 504 GB/s | ~30 tok/s | ~34 tok/s | ~53 tok/s |
| RTX 5070 | 672 GB/s | ~40 tok/s | ~46 tok/s | ~71 tok/s |
| RTX 3080 Ti | 912 GB/s | ~54 tok/s | ~62 tok/s | ~96 tok/s |
All are estimates from our calculator. NVIDIA cards work with every tool. AMD and Intel work well with Ollama and llama.cpp-based apps, with less support elsewhere.
The RTX 3060 12 GB, by the numbers
What doesn't fit
gpt-oss 20B needs about 16 GB, and the 27B to 32B models need 19 to 22 GB. On a 12 GB card they spill onto the CPU and slow to a few tokens per second. A 3-bit squeeze of a bigger model is possible, but Qwen3 14B at 4-bit is usually the better trade.
Tips for 12 GB
- Watch context with the 14B. Pasting a long document can push it over 12 GB. If speed suddenly drops, run
ollama psand check it still says100% GPU. If not, switch to Gemma 3 12B or Qwen3 8B for long documents. - Close GPU-heavy apps. Games and some browser tabs hold VRAM.
- Mind the 3060 8 GB. There is also an 8 GB RTX 3060 with less bandwidth. Make sure a used one is the 12 GB model.
Why 12 GB is a good starting point
The cheapest real way into local AI is our "Starter" build: a used desktop with an RTX 3060 12 GB, about $300 to $400 all in, with the card itself around $180 to $230. Parts list on the builds page.
When you outgrow it, the next useful step is 16 GB for gpt-oss 20B, or 24 GB for the 30B class. For code specifically, see the best local LLM for coding.
One email when there's a new guide, podcast episode, or result from the rack. No spam.