Model profile
Mixtral 8x22B Instruct
Evidence summary
Safety. Mixtral 8x22B Instruct has an estimated Safety rank of #304; its 90% source-sensitivity interval is #43–#326. Its behavior-only rank is #296; company governance moves the combined estimate to #304. Published Safety evidence spans 9 eval lineages and 5 of 7 components. Its strongest relative result is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #72 of 248); its weakest is Qwen2 Safety Panel (illegal, #3 of 3).
Freedom. Mixtral 8x22B Instruct has an estimated Freedom rank of #83; its 90% source-sensitivity interval is #35–#348. Published Freedom evidence spans 8 eval lineages and 1 of 1 components. Its strongest relative result is HELM Safety (simple_safety_tests, #7 of 80); 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
- OpenRouter ↗mistralai/mixtral-8x22b-instruct
- Official model documentation ↗Family-level model document · mistral · first party
- Release source ↗direct research preview identity date
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 | #71 / 80 | ↑0.44 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #195 / 248 | ↑11.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #72 / 248 | ↑90.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #177 / 248 | ↑53.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #90 / 246 | ↑97.45 | Source ↗official | |
| FORTRESSaverage_risk_score | #42 / 50 | ↓56.06 | Source ↗official | |
| FORTRESSover_refusal_score | #17 / 49 | ↓2.77 | Source ↗official | |
| HELM Safetyanthropic_red_team | #70 / 80 | ↑0.962 | Source ↗official | |
| HELM Safetybbq | #52 / 80 | ↑0.916 | Source ↗official | |
| HELM Safetyharmbench | #68 / 80 | ↑0.521 | Source ↗official | |
| HELM Safetysimple_safety_tests | #74 / 80 | ↑0.915 | Source ↗official | |
| HELM Safetyxstest | #55 / 80 | ↑0.941 | Source ↗official | |
| Qwen2 Safety Panelfraud | #3 / 3 | ↓8.49 | Source ↗official | |
| Qwen2 Safety Panelillegal | #3 / 3 | ↓6.87 | Source ↗official | |
| Qwen2 Safety Panelpornography | #3 / 3 | ↓33.82 | Source ↗official | |
| Qwen2 Safety Panelprivacy | #3 / 3 | ↓15.03 | 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 | #10 / 80 | ↓0.44 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #175 / 248 | ↓90.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #71 / 248 | ↓53.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #154 / 246 | ↓97.45 | Source ↗official | |
| FORTRESSaverage_risk_score | #9 / 50 | ↑56.06 | Source ↗official | |
| FORTRESSover_refusal_score | #17 / 49 | ↓2.77 | Source ↗official | |
| HELM Safetyanthropic_red_team | #11 / 80 | ↓0.962 | Source ↗official | |
| HELM Safetyharmbench | #13 / 80 | ↓0.521 | Source ↗official | |
| HELM Safetysimple_safety_tests | #7 / 80 | ↓0.915 | Source ↗official | |
| HELM Safetyxstest | #55 / 80 | ↑0.941 | Source ↗official | |
| Qwen2 Safety Panelfraud | #1 / 3 | ↑8.49 | Source ↗official | |
| Qwen2 Safety Panelillegal | #1 / 3 | ↑6.87 | Source ↗official | |
| Qwen2 Safety Panelpornography | #1 / 3 | ↑33.82 | Source ↗official | |
| Qwen2 Safety Panelprivacy | #1 / 3 | ↑15.03 | Source ↗official |