Model profile
Phi 4
Evidence summary
Safety. Phi 4 has an estimated Safety rank of #156; its 90% source-sensitivity interval is #16–#278. Its behavior-only rank is #158; company governance moves the combined estimate to #156. Published Safety evidence spans 6 eval lineages and 5 of 7 components. Its strongest relative result is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #20 of 248); its weakest is Confabulations (confabulation_rate, #47 of 52).
Freedom. Phi 4 has an estimated Freedom rank of #557; its 90% source-sensitivity interval is #324–#610. Published Freedom evidence spans 5 eval lineages and 1 of 1 components. Its strongest relative result is UGI Leaderboard — base-model willingness (willingness_direct_score, #21 of 156); its weakest is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #229 of 248).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗phi-4
- Hugging Face ↗microsoft/phi-4
- OpenRouter ↗microsoft/phi-4
- 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.
- nguyenthilaitrieulong/phi-4-abliterated ↗Direct non-quantized finetune documented by Hugging Face metadata
- trohrbaugh/Phi-4-reasoning-vision-15B-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 |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #175 / 345 | ↓0.8119 | Source ↗official | |
| AILuminate General Purpose AI Chatchild_sexual_exploitation | #7 / 32 | ↓10.3 | Source ↗official | |
| AILuminate General Purpose AI Chatdefamation | #4 / 32 | ↓3.5 | Source ↗official | |
| AILuminate General Purpose AI Chathate | #4 / 32 | ↓1.9 | Source ↗official | |
| AILuminate General Purpose AI Chatindiscriminate_weapons_cbrne | #6 / 32 | ↓8.7 | Source ↗official | |
| AILuminate General Purpose AI Chatintellectual_property | #4 / 32 | ↓3 | Source ↗official | |
| AILuminate General Purpose AI Chatnon_violent_crimes | #5 / 32 | ↓5.1 | Source ↗official | |
| AILuminate General Purpose AI Chatprivacy | #4 / 32 | ↓2.5 | Source ↗official | |
| AILuminate General Purpose AI Chatsex_related_crimes | #4 / 32 | ↓5.2 | Source ↗official | |
| AILuminate General Purpose AI Chatsexual_content | #7 / 32 | ↓8.9 | Source ↗official | |
| AILuminate General Purpose AI Chatspecialized_advice | #5 / 32 | ↓8 | Source ↗official | |
| AILuminate General Purpose AI Chatsuicide_self_harm | #7 / 32 | ↓7.1 | Source ↗official | |
| AILuminate General Purpose AI Chatviolent_crimes | #4 / 32 | ↓5.9 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #50 / 104 | ↓35.89 | Source ↗official | |
| Confabulationsconfabulation_rate | #47 / 52 | ↓52.48 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #111 / 248 | ↑16.8 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #20 / 248 | ↑95 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #29 / 248 | ↑96.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #34 / 246 | ↑99.18 | Source ↗official | |
| HalluVerse-M3 Hallucination Recognitionhallucination_recognition_accuracy | #8 / 14 | ↑70.38 | 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 |
|---|---|---|---|---|
| AILuminate General Purpose AI Chatchild_sexual_exploitation | #26 / 32 | ↑10.3 | Source ↗official | |
| AILuminate General Purpose AI Chatdefamation | #29 / 32 | ↑3.5 | Source ↗official | |
| AILuminate General Purpose AI Chathate | #29 / 32 | ↑1.9 | Source ↗official | |
| AILuminate General Purpose AI Chatindiscriminate_weapons_cbrne | #27 / 32 | ↑8.7 | Source ↗official | |
| AILuminate General Purpose AI Chatintellectual_property | #29 / 32 | ↑3 | Source ↗official | |
| AILuminate General Purpose AI Chatnon_violent_crimes | #28 / 32 | ↑5.1 | Source ↗official | |
| AILuminate General Purpose AI Chatprivacy | #29 / 32 | ↑2.5 | Source ↗official | |
| AILuminate General Purpose AI Chatsex_related_crimes | #29 / 32 | ↑5.2 | Source ↗official | |
| AILuminate General Purpose AI Chatsexual_content | #25 / 32 | ↑8.9 | Source ↗official | |
| AILuminate General Purpose AI Chatspecialized_advice | #27 / 32 | ↑8 | Source ↗official | |
| AILuminate General Purpose AI Chatsuicide_self_harm | #26 / 32 | ↑7.1 | Source ↗official | |
| AILuminate General Purpose AI Chatviolent_crimes | #29 / 32 | ↑5.9 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #55 / 104 | ↑35.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #229 / 248 | ↓95 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #220 / 248 | ↓96.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #212 / 246 | ↓99.18 | Source ↗official | |
| SpeechMap model completioncomplete_pct | #110 / 181 | ↑49.2 | Source ↗official | |
| UGI Leaderboard — base-model willingnesswillingness_adherence_score | #126 / 156 | ↑1 | 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.