Yes — with 9.6 GB to spare
Gemma 4 26B-A4B at Q4_K_M fits your CPU only · DDR4 dual-channel entirely on the GPU at 8K context, at an estimated 4.4 tokens per second. Past 253K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Mixture of experts with 3.8B active. Slower to think than Qwen3.6 35B-A3B, faster to answer, and it sees images.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 26.2 GB | 27.3 GB | — | ~2.5 | −0.1% ppl | 1.7 GB over |
| Q6_K | 20.2 GB | 21.3 GB | 116K | 3.3 | −0.4% ppl | Long context |
| Q5_K_M | 17.5 GB | 18.6 GB | 187K | 3.8 | −0.8% ppl | Long context |
| Q4_K_M | 14.9 GB | 16.0 GB | 253K | 4.4 | −1.9% ppl | Recommended |
| Q3_K_M | 12.1 GB | 13.2 GB | 256K | 5.5 | −5.4% ppl | Long context |
| Q2_K | 10.3 GB | 11.4 GB | 256K | 6.4 | −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, 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 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.