Can an Arc B580 run Gemma 4 31B Instruct?

In practice

It does not fit entirely. About 30 of the 60 layers of Gemma 4 31B Instruct fit on the Arc B580 at Q4_K_M, and the rest run on the CPU from system RAM. The estimate is about 2.6 tokens per second.

That number is lower than the share on the card suggests, and our own offload sweep shows why: every token waits for the slowest part, so the last layers left in RAM cost far more than the first ones moved. There is no quantization of this model worth using that fits entirely, so the choice is between this split and a smaller model.

The numbers

Gemma 4 31B Instruct has 31.27 billion parameters across 60 layers. The Arc B580 has 12 GB of VRAM at 456 GB/s, of which about 11.2 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.

The settings we would use

llama-server -m <model>.gguf -ngl 30 -c 8192 -fa on

30 of 60 layers fit on the card. If the load fails with an out-of-memory error, lower the number by 2 and try again.

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 Arc B580 at a quantization worth using:

• Gemma 3 12B Instruct, 12.19B at Q6_K, 10.52 GB, about 25.6 tokens/s
• Gemma 4 12B Instruct, 11.96B at Q6_K, 10.84 GB, about 24.8 tokens/s
• Gemma 2 9B Instruct, 9.24B at Q6_K, 9.42 GB, about 28.7 tokens/s

Shortening the context will not rescue this one: at 4k it still wants 20.06 GB against 21.31 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 Gemma 4 31B Instruct at a quantization worth using is the GeForce RTX 3090, around $700 used , running it at Q4_K_M at about 33.0 tokens per second. Check the current price (affiliate link, indicative price reviewed 2026-09-01). Compare every option.

What else an Arc B580 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 Gemma 4 31B Instruct 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 at2.6 tok/s

Whole answer1 min 7 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 Arc B580 runs, the full breakdown for Gemma 4 31B Instruct, how the numbers are made in the methodology, or a different pairing in the calculator.