Gemma 3 27B IT in Google’s quantization-aware Q4_0 GGUF (17,229,630,496 bytes) is 17.20 to 17.97 GiB at 8,192 tokens and 19.08 to 19.98 GiB at 32,768 tokens. KV moves from 1.03 GiB to 2.91 GiB. A full-attention cache with the same per-token cost at the short length would quadruple. It does not, because 10 layers keep the full context and 52 layers keep a 1,024-token window. Formula estimates, calculator 0.3.0, dataset 2026-10-02.1, llama.cpp, 1 sequence, FP16 KV, vision off, typical extras. Software support is unverified. Tokens per second are Unknown.
The files are Google’s QAT GGUF repos (google/gemma-3-*-it-qat-q4_0-gguf), not a Q4_K_M substitute. This catalog has no Q5 or Q8 Gemma file.
Four sizes, text only
| Model | File bytes | Total at 1,024 | KV | Total at 8,192 | KV | Total at 32,768 | KV |
|---|---|---|---|---|---|---|---|
| Gemma 3 1B IT | 1,003,541,152 | 1.07 to 1.47 | 0.01 | 1.10 to 1.50 | 0.04 | 1.20 to 1.60 | 0.14 |
| Gemma 3 4B IT | 3,155,051,328 | 3.20 to 3.63 | 0.13 | 3.33 to 3.79 | 0.27 | 3.80 to 4.29 | 0.74 |
| Gemma 3 12B IT | 8,074,473,920 | 8.02 to 8.55 | 0.38 | 8.46 to 9.04 | 0.81 | 9.96 to 10.64 | 2.31 |
| Gemma 3 27B IT | 17,229,630,496 | 16.66 to 17.36 | 0.48 | 17.20 to 17.97 | 1.03 | 19.08 to 19.98 | 2.91 |
No device is selected in that table, so the status is demand-only. The 8,192 to 32,768 KV ratios are about 3.5×, 2.7×, 2.9× and 2.8×. Qwen3-8B over the same step is exactly 4× (1.13 GiB to 4.50 GiB) on Qwen3 VRAM by size.
Gemma 3 27B on an RTX 4090
Net budget 22.00 GiB. Vision off.
| Tokens | KV GiB | Total GiB | Verdict |
|---|---|---|---|
| 1,024 | 0.48 | 16.66 to 17.36 | likely-fits |
| 8,192 | 1.03 | 17.20 to 17.97 | likely-fits |
| 32,768 | 2.91 | 19.08 to 19.98 | tight |
| 131,072 | 10.41 | 26.58 to 28.00 | insufficient |
131,072 is the native context on the 4B, 12B and 27B configs. The 1B native context is 32,768, and the table stops there for that size.
The catalog records a separate vision component of 857,739,168 bytes on the 27B Q4_0 artifact (Google’s mmproj file size stored with the GGUF). With vision included, 8,192 tokens, RTX 4090, the total is 18.00 to 18.77 GiB, likely-fits, coverage full. The catalog note says each image adds 256 tokens per the model card. This page does not add a further image-activation measurement. That number is Unknown.
Layer splits and the longer curves are on KV cache size by model. Model pages: Gemma 3 1B, Gemma 3 4B, Gemma 3 12B, Gemma 3 27B.
FAQ
How much VRAM does Gemma 3 27B Q4 need?
Google's QAT GGUF Q4_0 is 17,229,630,496 bytes. At 8,192 text tokens, llama.cpp, FP16 KV, vision off, the total is 17.20 to 17.97 GiB. At 32,768 tokens it is 19.08 to 19.98 GiB.
Why doesn't KV quadruple from 8K to 32K?
Gemma 3 27B keeps a full-length cache in 10 of 62 layers and a 1,024-token window in the other 52. KV moves from 1.03 GiB to 2.91 GiB. Qwen3-8B, which caches every layer in full, moves from 1.13 GiB to 4.50 GiB over the same step.
Does the text total include the image projector?
The rows below leave vision off. The 27B catalog component is 857,739,168 bytes. Turning it on at 8,192 tokens on an RTX 4090 makes the total 18.00 to 18.77 GiB. A measured image activation is Unknown.