Only with CPU offload

Gemma 4 26B-A4B at Q4_K_M needs 16.0 GB but only 14.4 GB is addressable, so about 11% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 31 tokens per second — usable for batch work, painful for chat.

89% on GPU 8K context Q4_K_M · 14.9 GB Apache 2.0 Released 2 Apr 2026 Vision

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

weights 14.9 GB

Weights 14.9 GB KV cache @ 8K 0.51 GB Runtime overhead 0.6 GB Over budget 1.6 GB past 14.4 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
Q8_0	26.2 GB	27.3 GB	—	~6.7	−0.1% ppl	12.9 GB over
Q6_K	20.2 GB	21.3 GB	—	~11	−0.4% ppl	6.9 GB over
Q5_K_M	17.5 GB	18.6 GB	—	~17	−0.8% ppl	4.2 GB over
Q4_K_M	14.9 GB	16.0 GB	—	~31	−1.9% ppl	1.6 GB over
Q3_K_M	12.1 GB	13.2 GB	39K	69	−5.4% ppl	Long context
Q2_K	10.3 GB	11.4 GB	83K	80	−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

Ollama llama.cpp LM Studio vLLM

terminal

$ llama-server \
    -hf google/gemma-4-26B-A4B-it:Q4_K_M \
    -c 8192 -ngl 26

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 14.9 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.

03If it falls back to CPU silently, drop the context first, then step down a quantisation.

See all models for this rig Compare with another model