Can an RTX 3090 run GPT OSS 120B?

In practice

GPT OSS 120B does not fit in 24 GB of VRAM on its own, and it does not have to. It is a mixture-of-experts model, so most of its weight is in experts of which only a few are read per token. Keeping the experts in system RAM and everything dense on the card is much faster than moving whole layers, and the engine puts about 38% of the experts on the GPU. The estimate is about 23.7 tokens per second.

Plan on at least 37 GB of free system RAM beyond what the card holds, and know that the speed is set mostly by how fast that RAM is. We have not measured a mixture-of-experts model on our own machine, so read the figure as good to about 30%. The command below is the llama.cpp way to set it up.

The numbers

GPT OSS 120B has 116.83 billion parameters across 36 layers, of which 5.7 billion are read for each token. The GeForce RTX 3090 has 24 GB of VRAM at 936 GB/s, of which about 23.2 GB is left for a model once the operating system has taken its share.

Nothing on this ladder fits entirely, including the 2-bit levels we do not recommend. Rows marked partial or experts in RAM run with part of the model in system RAM, at the speed shown. This model is published in a single native format, so every row is the same file.

The settings we would use

llama-server -m <model>.gguf -ngl all --n-cpu-moe 22 -c 8192 -fa on

The first 22 of 36 layers keep their experts in system RAM, and everything dense stays on the GPU. Lower the number until the load fails, then go back one.

This split is set in llama.cpp. Ollama chooses its own placement and does not take these options, so run ollama ps to see what it picked. The llama.cpp flags guide explains each option above.

If you want it faster

A smaller model from the same family is the straightforward answer. These fit on an RTX 3090 at a quantization worth using:

• GPT OSS 20B, 20.91B at MXFP4, 11.55 GB, about 167 tokens/s

Shortening the context will not rescue this one: at 4k it still wants 59.26 GB against 59.4 GB at 8k, because the weights, not the cache, are what fill the card.

Models of this size are what a second GPU or a large-memory machine is for. The multi-GPU guide shows what 2 cards add, and the Mac and PC comparison shows where unified memory holds a model that no single card can.

Or the hardware that does run it

The cheapest device we track that runs GPT OSS 120B at a quantization worth using is the GeForce GTX 1660 SUPER, around $110 used , running it at MXFP4 at about 15.4 tokens per second. Check the current price (affiliate link, indicative price reviewed 2026-09-01). Compare every option.

What else an RTX 3090 runs well

For comparison, these are the other popular models that fit this device at a sound quantization, best pairings first.

The same model on nearby hardware

Memory decides whether GPT OSS 120B loads and bandwidth decides how quickly it answers. Here is what a step down, a card with the same memory and a step up do with it.

Wait for the first word350 ms

Then writes at23.7 tok/s

Whole answer7.6 s

YouWhy does my model use more memory when the conversation gets longer?

Model

Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.

Questions people ask

Keep going

Everything the GeForce RTX 3090 runs, the full breakdown for GPT OSS 120B, how the numbers are made in the methodology, or a different pairing in the calculator.