Model profile
Mistral Small 3
Evidence summary
Safety. Mistral Small 3 has an estimated Safety rank of #268; its 90% source-sensitivity interval is #39–#325. Its behavior-only rank is #256; company governance moves the combined estimate to #268. Published Safety evidence spans 9 eval lineages and 6 of 7 components. Its strongest relative result is Vectara HHEM Factual Consistency (factual_consistency_rate, #8 of 94); its weakest is HELM Safety (harmbench, #76 of 80).
Freedom. Mistral Small 3 has an estimated Freedom rank of #18; its 90% source-sensitivity interval is #13–#296. Published Freedom evidence spans 7 eval lineages and 1 of 1 components. Its strongest relative result is UGI Leaderboard — base-model willingness (willingness_adherence_score, #6 of 156); its weakest is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #175 of 248).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗mistral-small-3
- Hugging Face ↗mistralai/Mistral-Small-24B-Instruct-2501
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗direct research preview identity date
Finetunes
Direct finetunes linked to this canonical base model. Quantized and repackaged derivatives are excluded.
- WasamiKirua/Sakura-24B-Spice ↗Direct non-quantized finetune documented by Hugging Face metadata
- Yingyaeliae/Dolphin-Mistral-24B-Venice-Edition-heretic ↗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 |
|---|---|---|---|---|
| AIRBench 2024 Safety Scenariossafety_scenarios | #76 / 80 | ↑0.328 | Source ↗official | |
| Confabulationsconfabulation_rate | #43 / 52 | ↓38.61 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #131 / 248 | ↑15.25 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #72 / 248 | ↑90.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #197 / 248 | ↑45.56 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #117 / 246 | ↑96.55 | Source ↗official | |
| HELM Safetyanthropic_red_team | #69 / 80 | ↑0.964 | Source ↗official | |
| HELM Safetybbq | #42 / 80 | ↑0.933 | Source ↗official | |
| HELM Safetyharmbench | #76 / 80 | ↑0.452 | Source ↗official | |
| HELM Safetysimple_safety_tests | #71 / 80 | ↑0.933 | Source ↗official | |
| HELM Safetyxstest | #48 / 80 | ↑0.953 | Source ↗official | |
| Vectara HHEM Factual Consistencyfactual_consistency_rate | #8 / 94 | ↑94.9 | 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.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AIRBench 2024 Safety Scenariossafety_scenarios | #5 / 80 | ↓0.328 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #175 / 248 | ↓90.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #51 / 248 | ↓45.56 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #130 / 246 | ↓96.55 | Source ↗official | |
| HELM Safetyanthropic_red_team | #12 / 80 | ↓0.964 | Source ↗official | |
| HELM Safetyharmbench | #5 / 80 | ↓0.452 | Source ↗official | |
| HELM Safetysimple_safety_tests | #10 / 80 | ↓0.933 | Source ↗official | |
| HELM Safetyxstest | #48 / 80 | ↑0.953 | Source ↗official | |
| UGI Leaderboard — base-model willingnesswillingness_adherence_score | #6 / 156 | ↑8 | Source ↗official | |
| UGI Leaderboard — base-model willingnesswillingness_direct_score | #21 / 156 | ↑5 | 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.