Short answer: yes. Gemma 3 27B at Q4_K_M needs about 18.9 GB at an 8K context, so it fits a 24 GB card with about 5 GB to spare. It does not fit 16 GB cards at Q4.
Check your own card with the Can My GPU Run This AI Model? calculator.
Why context is cheap on Gemma 3
Gemma 3 27B has 62 layers, but only every sixth layer (10 of them) uses full attention. The other 52 use a 1,024-token sliding window. As a result the KV cache grows slowly with context.
| Context | KV cache | Total (est., Q4_K_M) |
|---|---|---|
| 8K | 1.11 GB | 18.9 GB |
| 16K | 1.78 GB | 19.6 GB |
| 32K | 3.12 GB | 20.9 GB |
| 64K | 5.80 GB | 23.6 GB |
Which GPUs can run it?
| GPU | VRAM | Needed (est.) | Result | Speed (est., tok/s) |
|---|---|---|---|---|
| RTX 5090 | 32 GB | 18.9 GB | Fits | 64 to 85 |
| RTX 4090 | 24 GB | 18.9 GB | Fits | 36 to 48 |
| RTX 3090 | 24 GB | 18.9 GB | Fits | 33 to 45 |
| RX 7900 XTX | 24 GB | 18.9 GB | Fits | 34 to 46 |
| RX 7900 XT | 20 GB | 18.9 GB | Tight | 29 to 38 |
What if your GPU has 16 GB?
You will need a smaller model: Gemma 3 12B (about 9.3 GB at Q4 with 8K context) or Qwen3 14B (about 11.4 GB) fit comfortably.
Frequently asked questions
Can an RTX 3090 run Gemma 3 27B?
Yes, about 19 GB at 8K context at Q4_K_M.
Can it run at Q8?
No. Q8 weights are about 29 GB, which does not fit on 24 GB.
Does a 20 GB card work?
Barely: the estimate is about 19 GB, which is tight on 20 GB.
Sources and method
- Architecture: Gemma 3 27B config on Hugging Face (unsloth mirror, because the Google repository requires a login). 62 layers, 16 KV heads of size 128, sliding window 1,024, sliding-window pattern 6.
Estimates: weights = parameters × bits per weight; KV cache from the model’s published architecture; about 1 GB of runtime overhead. “Fits” means at most 92 percent of VRAM, “Tight” is 92 to 100 percent. Speed is memory bandwidth divided by model size at 60 to 80 percent efficiency. These are planning estimates, not our own benchmarks.
Last updated 3 Nov 2026.
