Which LLMs Fit in 12 GB of VRAM?

12 GB is enough for the 12B to 14B class at Q4_K_M. Gemma 4 12B ranks first and needs 8.6 GB, which leaves room for a longer context. Qwen3 14B is the largest that fits, at 11.4 GB, and Phi-4 14B needs 11.3 GB. Both are tight, with little left for the desktop.

Below those, every 7B to 9B model in the calculator fits with room to spare. The order follows the model finder's best pick, which weighs size against age. Use it to pick a few to try, not as a final verdict.

At 32K, the 14B models drop out. Qwen3 14B grows from 11.4 GB to 15.9 GB, while Gemma 4 12B only goes from 8.6 GB to 9.1 GB. Gemma 4 uses sliding window attention on most of its layers, and those layers only cache the most recent tokens.

For long chats or big documents on 12 GB, start from the 32K columns.

At Q8_0 with 8K, most 8B and 9B models fit. Llama 3.1 8B needs 10.6 GB, and Qwen3.5 9B is the largest at 11.6 GB, which is tight. At 32K they no longer fit, and the largest model left at Q8_0 is Gemma 4 E4B, at 10.0 GB.

So on 12 GB, the choice is often an 8B model at Q8_0 or a 12B to 14B model at Q4_K_M. The bigger model at Q4_K_M usually wins on quality. Q8_0 makes more sense when the smaller model already handles your task well.

gpt-oss 20B needs 13.7 GB at Q4_K_M with 8K, 1.7 GB too much. It's a mixture of experts model, so only a few of its experts run for each token. That makes it a good fit for a split: runners like llama.cpp can keep some expert weights in system RAM and the rest on the card, and the speed loss is smaller than with a dense model.

The 24B to 26B MoE models are further out: LFM2.5 24B-A2B needs 16.3 GB and Gemma 4 26B (MoE A4B) 17.4 GB. The 24 GB list covers them.

Nvidia's RTX 5070, RTX 4070, RTX 4070 Super and RTX 3060 12GB all have 12 GB, and so do Intel's Arc B580 and AMD's RX 7700 XT. They run the same list. What changes is speed: the RTX 3060 has about half the memory bandwidth of an RTX 5070, so it writes tokens more slowly.

Nvidia also sold an 8 GB RTX 3060, which belongs on the 8 GB list. If the card drives your monitors, keep some memory free for the desktop.