Model profile
Gemma 4 31B It
Evidence summary
Gemma 4 31B It has an estimated overall rank of #115; its 90% source-sensitivity interval is #50–#194. Its behavior-only rank is #125; company governance moves the combined estimate to #115. Published evidence spans 8 evals and 7 of 7 behavior components. Its strongest relative result is SM-Bench (adversarial, #2 of 73); its weakest is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #253 of 260).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gemma-4-31b
- Hugging Face ↗google/gemma-4-31B-it
- NVIDIA ↗qc69jvmznzxy/gemma-4-31b-it
- OpenRouter ↗google/gemma-4-31b-it
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗repository created at proxy
- huggingface.co ↗
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 | #166 / 311 | 0.8158 | ↓ lower | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #59 / 105 | 0.225 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #237 / 260 | 8.53 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #253 / 260 | 66.67 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #40 / 260 | 93.89 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #147 / 258 | 95.64 | ↑ higher | Source ↗official | |
| PHAREbias_resistance_diagnostic | #54 / 66 | 0.3522 | ↑ higher | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #30 / 70 | 0.7662 | ↑ higher | Source ↗official | |
| PHAREharm_resistance_diagnostic | #17 / 70 | 0.9651 | ↑ higher | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #24 / 67 | 0.6122 | ↑ higher | Source ↗official | |
| SM-Benchadversarial | #2 / 73 | 91.71 | ↑ higher | Source ↗official | |
| SM-Benchambiguous_interpretation | #9 / 73 | 90.77 | ↑ higher | Source ↗official | |
| SM-Benchanti_hallucination | #19 / 73 | 96.86 | ↑ higher | Source ↗official | |
| SM-Bencheq_boundaries | #26 / 73 | 67.7 | ↑ higher | Source ↗official | |
| SM-Benchoverfit | #8 / 73 | 91.8 | ↑ higher | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #16 / 102 | 6.98 | ↑ higher | Source ↗official | |
| SpeciEvalland_animal_4ns | #42 / 102 | 4.47 | ↓ lower | Source ↗official | |
| SpeciEvalsea_animal_4ns | #63 / 102 | 4.85 | ↓ lower | Source ↗official | |
| SpeciEvalspeciesism | #94 / 102 | 2.88 | ↓ lower | Source ↗official | |
| TACbase_welfare_rate | #47 / 68 | 23.72 | ↑ higher | Source ↗self run | |
| Vigil Mental Health Safetyoverall_score | #12 / 23 | 46 | ↑ 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.