Model profile
Gemma 4 31B IT
Evidence summary
Safety. Gemma 4 31B IT has an estimated Safety rank of #79; its 90% source-sensitivity interval is #64–#179. Its behavior-only rank is #91; company governance moves the combined estimate to #79. Published Safety evidence spans 17 eval lineages and 7 of 7 components. Its strongest relative result is SM-Bench (adversarial, #2 of 84); its weakest is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #234 of 248).
Freedom. Gemma 4 31B IT has an estimated Freedom rank of #362; its 90% source-sensitivity interval is #191–#496. Published Freedom evidence spans 7 eval lineages and 1 of 1 components. Its strongest relative result is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #15 of 248); its weakest is SM-Bench (adversarial, #82 of 84).
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 ↗
Finetunes
Direct finetunes linked to this canonical base model. Quantized and repackaged derivatives are excluded.
huihui-ai/Huihui-gemma-4-31B-it-abliterated ↗Direct non-quantized finetune documented by Hugging Face metadata- huihui-ai/Huihui-gemma-4-31B-it-abliterated-v2 ↗Direct non-quantized finetune documented by Hugging Face metadata
- llmfan46/gemma-4-31B-it-uncensored-heretic ↗Direct non-quantized finetune documented by Hugging Face metadata
- wangzhang/gemma-4-31B-it-abliterated ↗Direct non-quantized finetune documented by Hugging Face metadata
Safety evals
Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #185 / 345 | ↓0.8195 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5child_sexual_exploitation | #19 / 31 | ↓51.6 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5defamation | #19 / 31 | ↓72.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5hate | #19 / 31 | ↓61.1 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5indiscriminate_weapons_cbrne | #17 / 31 | ↓73.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5intellectual_property | #19 / 31 | ↓67.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5non_violent_crimes | #17 / 31 | ↓82.1 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5privacy | #19 / 31 | ↓72.2 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5sex_related_crimes | #21 / 31 | ↓78.8 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5sexual_content | #20 / 31 | ↓55.2 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5specialized_advice | #19 / 31 | ↓68 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5suicide_self_harm | #18 / 31 | ↓61.8 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5violent_crimes | #18 / 31 | ↓75.8 | Source ↗official | |
| ANIMAscore | #15 / 22 | ↑0.6863 | Source ↗self run | |
| Arena Factuality — Text Arena (factuality-only weighting)factuality_bt_rating | #43 / 111 | ↑1442.0 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #71 / 117 | ↑0.225 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #226 / 248 | ↑8.53 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #234 / 248 | ↑66.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #38 / 248 | ↑93.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #140 / 246 | ↑95.64 | Source ↗official | |
| JuICE Cultural-Error Span Detectionf1 | #5 / 10 | ↑0.445 | Source ↗official | |
| KIDBench Implicit Child Cueimplicit_child_cue_total_mean | #4 / 13 | ↑4.14 | Source ↗official | |
| Opposite-Narrator Sycophancysycophancy_rate_pct | #16 / 24 | ↓4.6 | Source ↗official | |
| PHAREbias_resistance_diagnostic | #54 / 66 | ↑0.3522 | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #30 / 70 | ↑0.7662 | Source ↗official | |
| PHAREharm_resistance_diagnostic | #17 / 70 | ↑0.9651 | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #24 / 67 | ↑0.6122 | Source ↗official | |
| RealityTest — Text AI-Identity Disclosuredisclosure_probability | #8 / 17 | ↑0.313 | Source ↗official | |
| SM-Benchadversarial | #2 / 84 | ↑91.71 | Source ↗official | |
| SM-Benchambiguous_interpretation | #12 / 84 | ↑90.77 | Source ↗official | |
| SM-Benchanti_hallucination | #26 / 84 | ↑96.86 | Source ↗official | |
| SM-Bencheq_boundaries | #31 / 84 | ↑67.7 | Source ↗official | |
| SM-Benchoverfit | #12 / 84 | ↑91.8 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #21 / 123 | ↑6.98 | Source ↗official | |
| SpeciEvalland_animal_4ns | #54 / 123 | ↓4.47 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #81 / 123 | ↓4.85 | Source ↗official | |
| SpeciEvalspeciesism | #115 / 123 | ↓2.88 | Source ↗official | |
| StereoTales Harmful Associationsbenign_significant_association_score | #5 / 23 | ↑88.1 | Source ↗official | |
| TACbase_welfare_rate | #58 / 87 | ↑23.72 | Source ↗self run | |
| Vectara HHEM Factual Consistencyfactual_consistency_rate | #31 / 94 | ↑92.6 | Source ↗official | |
| Vigil Mental Health Safetyoverall_score | #12 / 23 | ↑46 | Source ↗official |
Freedom evals
Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.
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.