Can an RTX 5070 run Gemma 4 31B Instruct?
No
Only barely. It loads at IQ2_XXS, which compresses the weights far enough to visibly degrade the model, so the honest answer is that this pairing does not work.
The numbers
Gemma 4 31B Instruct has 31.27 billion parameters across 60 layers. The GeForce RTX 5070 has 12 GB of VRAM at 672 GB/s, of which about 11.2 GB is left for a model once the operating system has taken its share.
| Quantisation | Total needed | Fits | Speed |
|---|---|---|---|
| Q8_0 | 32.81 GB | no | — |
| Q6_K | 25.75 GB | no | — |
| Q5_K_M | 22.58 GB | no | — |
| Q4_K_M | 19.45 GB | no | — |
| IQ4_XS | 17.34 GB | no | — |
| Q3_K_M | 16.1 GB | no | — |
| IQ3_XXS degraded | 13.01 GB | no | — |
| Q2_K degraded | 14.06 GB | no | — |
| IQ2_XXS degraded | 9.37 GB | yes | 65.3 t/s |
1 of the 9 levels fit. Levels marked degraded are listed for completeness, not as advice.
What to do instead
The straightforward answer is a smaller model from the same family. These do fit on a GeForce RTX 5070, at a quantisation worth using:
- vLLM Translategemma 12B Instruct — 13.19B at Q5_K_M, 9.95 GB, about 60.5 tokens/s
- Gemma 3 12B Instruct — 12.19B at Q6_K, 10.96 GB, about 54.4 tokens/s
- Gemma 4 12B Instruct — 11.96B at Q6_K, 10.34 GB, about 58 tokens/s
Shortening the context window will not rescue this one: at 4k it still wants 19.29 GB against 19.45 GB at 8k, because the weights rather than the cache are what fill the card. The gap here is too large for settings to close.
Or the hardware that does run it
The cheapest device we track that runs Gemma 4 31B Instruct at a quantisation worth using is the Radeon RX 7900 XT, around $650 , running it at IQ4_XS at about 35.1 tokens per second. Check the current price (affiliate link; indicative price reviewed 2026-09-01). Compare every option.
See how fast it feels
GeForce RTX 5070 running Gemma 4 31B Instruct at IQ2_XXS
YouWhy does my model use more memory when the conversation gets longer?
Because of the KV cache. Every token you send leaves behind a key and a value vector in each layer of the model, and those stay in memory for as long as the conversation lasts.
The weights are a fixed cost: load a 4-bit 8B model and that is about 4.8 GB, whether you write one word or ten thousand. The cache is the part that grows, and it grows in a straight line with the number of tokens in the window.
How steeply depends on the model's attention design. With grouped-query attention, several query heads share one key-value pair, which cuts the cache by that ratio. Without it, every head keeps its own, and a long context can cost more memory than the weights themselves.
If you are short on memory, the first thing to try is lowering the context window in your runtime, not the quantisation.
Reading your question
Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU. The same answer on RTX 4090 at Q5_K_M would take about 4.8 s.
Keep going
Everything the GeForce RTX 5070 runs, the full breakdown for Gemma 4 31B Instruct, or check a different pairing in the calculator.