Only with CPU offload
Gemma 4 26B-A4B at Q4_K_M needs 16.0 GB but only 10.6 GB is addressable, so about 36% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 14 tokens per second — usable for batch work, painful for chat.
Mixture of experts with 3.8B active. Slower to think than Qwen3.6 35B-A3B, faster to answer, and it sees images.
What hardware do I need for Gemma 4 26B-A4B? →
Fits instead: Gemma 4 12B (8.2 GB) · Gemma 4 E4B (5.2 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 26.2 GB | 27.3 GB | — | ~5.3 | −0.1% ppl | 16.7 GB over |
| Q6_K | 20.2 GB | 21.3 GB | — | ~8.0 | −0.4% ppl | 10.7 GB over |
| Q5_K_M | 17.5 GB | 18.6 GB | — | ~10 | −0.8% ppl | 8.0 GB over |
| Q4_K_M | 14.9 GB | 16.0 GB | — | ~14 | −1.9% ppl | 5.4 GB over |
| Q3_K_M | 12.1 GB | 13.2 GB | — | ~24 | −5.4% ppl | 2.6 GB over |
| Q2_K | 10.3 GB | 11.4 GB | — | ~42 | −15% ppl | 0.8 GB over |
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, 5 of 30 layers global), which is why its cache barely grows with context.
How to run it
$ llama-server \
-hf google/gemma-4-26B-A4B-it:Q4_K_M \
-c 8192 -ngl 19
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.