‹ PublicAI Index
The LLM benchmark aggregator.
Gemma 4 26B A4B
Google · 26B · A4B · open weights
Strongest in Factual grounding (#14 of 104), weakest in Agents (#237 of 298). Above par in 6 of 13 scopes. Among the models it meets almost everywhere, it finishes behind Claude Sonnet 5 and Claude Opus 5 and ahead of MiniMax M3 and Kimi K2.
§ 1 · Profile
What it is good at
Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.
Human preference60.2−8.6#85/3541/1
Human preference60.2−8.6#85/3541/1
Knowledge49.5−13#119/1971/3
Document parsing49.5−13.8#50/831/1
Safety50.4−11#165/3272/3
Factual grounding62−8.7#14/1041/1
Secure code61−5.2#55/2681/1
Toxicity avoidance54.7−4.3#110/2661/1
Harm refusal47.6−13.9#195/2901/2
Fairness44−26#210/2901/2
Jailbreak resistance39.2−25.8#217/2661/1
Agents43.7−22.6#237/2982/6
Knowledge work40−30.7#152/2082/2
§ 2 · Head to head
What it beats, and what beats it
The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for Gemma 4 26B A4B, left for the other.
§ 3 · Sources
Where the numbers come from
6 publications, 16 figures. Every one links to the page it was read from.
LMArena Text 1437
GDPval-AA 573
AA-Briefcase 316
ParseBench · Overall 58.5%ParseBench · Tables 70%ParseBench · Charts 14.2%ParseBench · Content faithfulness 83.8%ParseBench · Semantic formatting 65.1%ParseBench · Visual grounding 59.2%
Enkrypt · Jailbreak risk 21.9%Enkrypt · Harmful content risk 5%Enkrypt · CBRN risk 37%Enkrypt · Toxicity risk 2.9%Enkrypt · Bias risk 87.1%Enkrypt · Insecure code risk 10.7%
Vectara · Factual consistency 94.8%
Badge
[](https://publicai.io/model-index/m/gemma-4-26b-a4b)