Only with CPU offload
Gemma 3 27B at Q4_K_M needs 17.0 GB but only 14.4 GB is addressable, so about 17% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 6.9 tokens per second — usable for batch work, painful for chat.
The 2025 single-GPU generalist with vision. Its Gemma-licence terms are the reason to prefer Gemma 4 now.
What hardware do I need for Gemma 3 27B? →
Fits instead: Gemma 3 12B (8.3 GB) · Gemma 3 4B (3.3 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 27.1 GB | 28.7 GB | — | ~2.1 | −0.1% ppl | 14.3 GB over |
| Q6_K | 20.9 GB | 22.6 GB | — | ~3.4 | −0.4% ppl | 8.2 GB over |
| Q5_K_M | 18.1 GB | 19.7 GB | — | ~4.6 | −0.8% ppl | 5.3 GB over |
| Q4_K_M | 15.4 GB | 17.0 GB | — | ~6.9 | −1.9% ppl | 2.6 GB over |
| Q3_K_M | 12.5 GB | 14.1 GB | 11K | 14 | −5.4% ppl | Fits |
| Q2_K | 10.7 GB | 12.3 GB | 34K | 16 | −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 51
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.