howtospark
models / gemma-4-26b-a4b-it

Gemma 4 26B-A4B

google/gemma-4-26B-A4B-it
params
26.5B
active params
4.0B
sparsity
MoE
attention
Full attention
architecture
gemma4
license
apache-2.0
downloads
14.4M
likes
1.3k
hf updated
Jul 10, 2026

Visualize

total size

each column is a layer — lit blocks are the experts that fire per token

bits / param
16-bit
footprint
49.4 GiB
sparks to fit
block ≈
207M params
active paramsdormant expertsghost = precision traded awayblock fill ∝ bits kept · Gemma 4 26B-A4B26.544131376B params

Eval scores (compare all)

Reported

77.1LiveCodeBench
88.3AIME 2026
82.6MMLU-Pro
82.3GPQA Diamond