Yes, just — 0.6 GB spare
Gemma 4 12B at Q4_K_M fits your GeForce RTX 3080 entirely on the GPU at 8K context, at an estimated 68 tokens per second. Past 18K 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 · 6.7 GB
Apache 2.0
Released 29 May 2026
Vision
The "unified" Gemma 4: text, image and audio in one 12B that fits a 12 GB card at Q4. 140+ languages.
The VRAM budget
weights 6.7 GB
Weights 6.7 GB
KV cache @ 8K 0.81 GB
Runtime overhead 0.6 GB
Free 0.6 GB of 8.8 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 22.4 GB | 23.8 GB | — | ~2.3 | Reference | 15.0 GB over |
| Q8_0 | 11.9 GB | 13.3 GB | — | ~7.2 | −0.1% ppl | 4.5 GB over |
| Q6_K | 9.2 GB | 10.6 GB | — | ~15 | −0.4% ppl | 1.8 GB over |
| Q5_K_M | 7.9 GB | 9.3 GB | — | ~33 | −0.8% ppl | 0.5 GB over |
| Q4_K_M | 6.7 GB | 8.2 GB | 18K | 68 | −1.9% ppl | Recommended |
| Q3_K_M | 5.5 GB | 6.9 GB | 38K | 84 | −5.4% ppl | Long context |
| Q2_K | 4.7 GB | 6.1 GB | 51K | 98 | −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, 8 of 48 layers global), which is why its cache barely grows with context.
How to run it
$ llama-server \
-hf google/gemma-4-12B-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 6.7 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 0.6 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 18K context.