What can you run with 12 GB of VRAM? 14B models, 20B experts and long context

12 GB is the first amount of VRAM where a local model stops feeling like a compromise. A 14B model fits at a good quantization, there is room for a real context window, and the models just above it become reachable with a trick or two. If you have the choice between 8 GB and 12 GB, this is the page that explains why we say take the 12.

What fits comfortably

The usable memory on a 12 GB card is about 11.2 GB, and the sweet spot is a 12B to 14B model at Q5 or Q6 with an 8k to 16k window. These are the strongest pairings on an RTX 3060 12 GB, which is the card most people have at this tier.

Pick	Model	Download	Memory	Speed
Best all-rounder	Gemma 4 12B Instruct	Q6_K	10.8 GB	25.4 t/s
Best for code	Qwen2.5 Coder 14B Instruct	Q4_K_M	9.81 GB	28.3 t/s
Best for reasoning	Qwen3 14B	Q5_K_M	11.0 GB	25.2 t/s
Largest that still runs well	Qwen3 14B	Q5_K_M	11.0 GB	25.2 t/s

Model	Parameters	Download	Memory	Speed	Longest context
Gemma 4 12B Instruct	11.96B	Q6_K	10.8 GB	25.4 t/s	8k
Qwen3.5 9B	9.65B	Q8_0	9.25 GB	30 t/s	64k
Qwen3 14B	14.77B	Q5_K_M	11.0 GB	25.2 t/s	8k
Gemma 3 12B Instruct	12.19B	Q6_K	10.5 GB	26.2 t/s	16k
Gemma 4 E4B Instruct	8B	Q8_0	8.42 GB	32.9 t/s	128k
Granite 4.1 8B	8.79B	Q8_0	10.3 GB	26.6 t/s	8k
Qwen2.5 14B Instruct	14.77B	Q4_K_M	9.81 GB	28.3 t/s	8k
Qwen3 8B	8.19B	Q8_0	9.02 GB	30.8 t/s	16k
Qwen2.5 Coder 14B Instruct	14.77B	Q4_K_M	9.81 GB	28.3 t/s	8k
Qwen2.5 VL 7B Instruct	8.29B	Q8_0	8.49 GB	32.7 t/s	32k
Phi 4	14.66B	Q4_K_M	9.96 GB	27.8 t/s	8k
Mistral Nemo Instruct 2407	12.25B	Q6_K	10.5 GB	26.2 t/s	8k

The difference from 8 GB is not only the larger models. It is that you stop choosing between quality and context. A 9B model at Q8 with a long window, a 14B at Q5 with a normal one, a 12B at Q6, all fit, and you pick by what you are doing. For reasoning work the 14B models are the best of them. For long documents, the smaller hybrid-attention models are, because their cache barely grows.

The 20B at the edge

The model everyone asks about at 12 GB is gpt-oss-20b. It is a mixture-of-experts model with 21B parameters and about 4B active, so it generates fast, and it ships in a native 4-bit format. The published file is 12.1 GB, which is 11.3 GiB, so the weights alone take nearly all of a 12 GB card. There is no room left for the cache and the runtime's own buffers, and a straight load fails or falls into slow offload.

The fix is the expert split. Keep a few layers of experts in system RAM with --n-cpu-moe in llama.cpp, put everything else on the card, and the model runs. Our engine estimates about 65.7 tokens per second on an RTX 3060 this way, with about 97% of the experts on the GPU. That is a good deal faster than any dense model of its quality, and it needs 32 GB of system RAM to be comfortable. The MoE guide covers how the split works.

12 GB is not one thing

The same memory comes at very different speeds. A used RTX 3080 12 GB has 912 GB/s and an RTX 3060 12 GB has 360, and generation follows bandwidth. Here are the 12 GB cards on a 14B model:

Device	Memory	Bandwidth	Qwen3 14B	Phi 4
GeForce RTX 3060 12GB	12 GB	360 GB/s	25.2 t/s	does not fit
GeForce RTX 4070	12 GB	504 GB/s	35.3 t/s	does not fit
GeForce RTX 5070	12 GB	672 GB/s	47.0 t/s	does not fit
GeForce RTX 3080 12GB	12 GB	912 GB/s	63.8 t/s	does not fit
Arc B580	12 GB	456 GB/s	24.6 t/s	does not fit
Radeon RX 7700 XT	12 GB	432 GB/s	25.6 t/s	does not fit

The used 3080 12 GB is the quiet bargain of this tier, if you can find one at a fair price. It has the memory of a 3060 and two and a half times the bandwidth, and it runs hot and draws 350 W, so it needs a good case and power supply. Intel's Arc B580 and AMD's RX 7700 XT are the other new options. They are cheaper and their software support is narrower, which is worth checking against the runtime you plan to use before you buy.

Which cards have 12 GB

Most 12 GB cards are older or mid-range, which is why the spread in speed is so wide. The list is sorted by bandwidth, so the rows at the top are the ones that answer fastest, and prices are the ones we last reviewed.

Device	Memory	Bandwidth	Price
GeForce RTX 3080 Ti Desktop GPU, used	12 GB	912 GB/s	$450
GeForce RTX 3080 12GB Desktop GPU, used	12 GB	912 GB/s	$400
GeForce RTX 5070 Desktop GPU	12 GB	672 GB/s	$560
GeForce RTX 4070 Ti Desktop GPU, used	12 GB	504 GB/s	$650
GeForce RTX 4070 SUPER Desktop GPU	12 GB	504 GB/s	$560
GeForce RTX 4070 Desktop GPU	12 GB	504 GB/s	$500
Arc B580 Desktop GPU	12 GB	456 GB/s	$260
GeForce RTX 4080 Laptop Laptop GPU	12 GB	432 GB/s	n/a
Radeon RX 7700 XT Desktop GPU	12 GB	432 GB/s	$400
Radeon RX 6700 XT Desktop GPU, used	12 GB	384 GB/s	$250
GeForce RTX 3060 12GB Desktop GPU, used	12 GB	360 GB/s	$200
GeForce RTX 2060 12GB Desktop GPU, used	12 GB	336 GB/s	$160

If you are comparing a pair of cards from this list, put the bandwidth difference next to the price difference. A card that costs 25% more and reads memory 2 times faster is usually the better buy, and the reverse is rarely true.

What 12 GB cannot do

It cannot hold a 27B to 32B model at a good quantization. These need around 16 to 20 GB at 4 bits, and on 12 GB they go into partial offload at a fraction of the speed. It also cannot run the largest video models comfortably, and image models above the SDXL class need quantized variants. For those, the next step is 16 GB, with 24 GB as the point where the models above 30B become comfortable, and the 16 GB guide shows what changes.

Our advice at this tier

If you are buying, get the widest 12 GB card you can find, and treat the cheap new narrow-bus ones with caution. If you already have one, run a 14B model at Q5 with an 8k to 16k window for chat and reasoning, and try the 20B expert split when you want speed. Check the pairing in the calculator before a long download, because at 12 GB the margins are small enough that half a gigabyte decides the answer.