Yes, just — 1.4 GB spare
Gemma 3 27B at Q4_K_M fits your RTX 4000 Ada entirely on the GPU at 8K context, at an estimated 14 tokens per second. Past 25K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully on GPU
8K context
Q4_K_M · 15.4 GB
Gemma Terms of Use
Released Mar 2025
Vision
The 2025 single-GPU generalist with vision. Its Gemma-licence terms are the reason to prefer Gemma 4 now.
The VRAM budget
weights 15.4 GB
Weights 15.4 GB
KV cache @ 8K 1.03 GB
Runtime overhead 0.6 GB
Free 1.4 GB of 18.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 27.1 GB | 28.7 GB | — | ~2.8 | −0.1% ppl | 10.3 GB over |
| Q6_K | 20.9 GB | 22.6 GB | — | ~5.2 | −0.4% ppl | 4.2 GB over |
| Q5_K_M | 18.1 GB | 19.7 GB | — | ~8.8 | −0.8% ppl | 1.3 GB over |
| Q4_K_M | 15.4 GB | 17.0 GB | 25K | 14 | −1.9% ppl | Recommended |
| Q3_K_M | 12.5 GB | 14.1 GB | 62K | 17 | −5.4% ppl | Long context |
| Q2_K | 10.7 GB | 12.3 GB | 85K | 20 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (1024 tokens, 1 global in 6), which is why its cache barely grows with context.
How to run it
$ llama-server \
-hf google/gemma-3-27b-it:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 15.4 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03Only 1.4 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 25K context.