Skip to content

Can you run Gemma 4 12B-it on RTX 3090?

Short answer

Yes — Gemma 4 12B-it fits in RTX 3090 with 16.3 GB of VRAM to spare.

Model file
6.6 GB
Needs in VRAM, with context
about 7.7 GB
Card memory
24 GB
Measured on this hardware
not measured

How this was worked out

Gemma 4 12B-it is a 6.6 GB file. Running it needs the weights in memory plus room for the context window and the runtime, which comes to roughly 7.7 GB in practice. RTX 3090 has 24 GB, which leaves 16.3 GB. That is arithmetic rather than a benchmark: it tells you whether the model loads, not how it feels.

What we have actually measured

We have not timed Gemma 4 12B-it on RTX 3090. The closest figure we have is 24 tokens per second on RTX 3060 12 GB, Vulkan backend. Rather than scale that number to your card and present the result as a benchmark, this page leaves it out. The method is published at /benchmarks if you would like to run it yourself.

If it is tight

There is room to spare, so the thing to watch is what else is using the card. Close anything doing 3D or video encoding before you judge the speed. Nightjar sizes the context window to the memory it finds, so a smaller card gets a shorter window rather than a failed load.

Smaller models that fit the same card

Related