Model profile
Gemma 4 26B A4B It
Evidence summary
Gemma 4 26B A4B It has an estimated overall rank of #116; its 90% source-sensitivity interval is #50–#203. Its behavior-only rank is #126; company governance moves the combined estimate to #116. Published evidence spans 5 evals and 5 of 7 behavior components. Its strongest relative result is Enkrypt AI Safety Leaderboard (harmful_attack_non_success_rate, #35 of 260); its weakest is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #256 of 260).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gemma-4-26b-a4b
- Hugging Face ↗google/gemma-4-26B-A4B-it
- OpenRouter ↗google/gemma-4-26b-a4b-it
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗repository created at proxy
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Better | Source |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #160 / 311 | 0.809 | ↓ lower | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #71 / 105 | 0.16 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #181 / 260 | 12.92 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #256 / 260 | 63 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #35 / 260 | 95 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #112 / 258 | 97.09 | ↑ higher | Source ↗official | |
| TACbase_welfare_rate | #31 / 68 | 28.21 | ↑ higher | Source ↗self run | |
| Vigil Mental Health Safetyoverall_score | #13 / 23 | 38 | ↑ higher | Source ↗official |
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.