Model profile
Kimi K3
Evidence summary
Safety. Kimi K3 has an estimated Safety rank of #43; its 90% source-sensitivity interval is #32–#115. Its behavior-only rank is #33; company governance moves the combined estimate to #43. Published Safety evidence spans 26 eval lineages and 7 of 7 components. Its strongest relative result is Manager Coercion Bench (fabrication_rate, #1 of 15); its weakest is Claude system cards — Gray Swan Q1+Q2 indirect prompt injection k=15 (attack_success_probability_k15_pct, #12 of 12).
Freedom. Kimi K3 has an estimated Freedom rank of #339; its 90% source-sensitivity interval is #144–#464. Published Freedom evidence spans 6 eval lineages and 1 of 1 components. Its strongest relative result is Human Pathogen Capabilities Test (HPCT) — overall refusal (refusal_rate_pct, #1 of 69); its weakest is SM-Bench (adversarial, #48 of 92).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗kimi-k3
- Hugging Face ↗moonshotai/Kimi-K3
- OpenRouter ↗moonshotai/kimi-k3
- Official model documentation ↗Family-level model document · moonshot ai · first party
Finetunes
Direct finetunes linked to this canonical base model. Quantized and repackaged derivatives are excluded.
- Uniboshi/Kimi-K3-Abliterated-V1 ↗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 | #95 / 358 | ↓0.532 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #10 / 122 | ↑0.73 | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #29 / 49 | ↓76.6 | Source ↗official | |
| CAIS Risk Indexbioweapons_assistance | #41 / 54 | ↓87.2 | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #15 / 55 | ↓44.7 | Source ↗official | |
| CAIS Risk Indexmachiavelli | #38 / 51 | ↓92.6 | Source ↗official | |
| CAIS Risk Indexmask | #12 / 57 | ↓9.4 | Source ↗official | |
| CAIS Risk Indexpolitical_manipulation | #27 / 51 | ↓47.6 | Source ↗official | |
| CAIS Risk Indextextquests_harm | #21 / 54 | ↓17 | Source ↗official | |
| CheatBench direct cheating propensitybiology_bioinformatics_cheating_rate_pct | #8 / 13 | ↓95 | Source ↗official | |
| CheatBench direct cheating propensityboard_games_cheating_rate_pct | #7 / 13 | ↓34.2 | Source ↗official | |
| CheatBench direct cheating propensitycreative_writing_cheating_rate_pct | #5 / 13 | ↓80 | Source ↗official | |
| CheatBench direct cheating propensityknowledge_work_cheating_rate_pct | #7 / 13 | ↓94.4 | Source ↗official | |
| CheatBench direct cheating propensitymathematical_research_cheating_rate_pct | #3 / 13 | ↓75 | Source ↗official | |
| CheatBench direct cheating propensitymenial_computation_cheating_rate_pct | #5 / 13 | ↓100 | Source ↗official | |
| CheatBench direct cheating propensitymultimodal_cheating_rate_pct | #7 / 13 | ↓93.3 | Source ↗official | |
| CheatBench direct cheating propensitysoftware_engineering_cheating_rate_pct | #11 / 13 | ↓52.6 | Source ↗official | |
| CheatBench direct cheating propensitysvg_competition_cheating_rate_pct | #7 / 13 | ↓55.6 | Source ↗official | |
| Claude system cards — Gray Swan Q1+Q2 indirect prompt injection k=15attack_success_probability_k15_pct | #12 / 12 | ↓52.7 | Source ↗official | |
| FORTRESSaverage_risk_score | #32 / 60 | ↓26.95 | Source ↗official | |
| FORTRESSover_refusal_score | #22 / 59 | ↓3.315 | Source ↗official | |
| Google Gemini 3.8 launch — Gray Swan indirect prompt injection k=15attack_success_probability_k15_pct | #14 / 15 | ↓52.7 | Source ↗official | |
| Human Pathogen Capabilities Test (HPCT) — overall refusalrefusal_rate_pct | #27 / 69 | ↑0 | Source ↗official | |
| kindbench v0.1.0 psychological safety rankingemotional_safety | #6 / 10 | ↑80.2 | Source ↗official | |
| kindbench v0.1.0 psychological safety rankingidentity_collapse | #4 / 10 | ↑88.8 | Source ↗official | |
| kindbench v0.1.0 psychological safety rankingsycophancy_spine | #1 / 10 | ↑100 | Source ↗official | |
| kindbench v0.1.0 psychological safety rankingvalue_integrity | #2 / 10 | ↑97.4 | Source ↗official | |
| Manager Coercion Benchcoercion_ladder_depth | #28 / 45 | ↓8.5 | Source ↗official | |
| Manager Coercion Benchfabrication_rate | #1 / 15 | ↓0 | Source ↗official | |
| Olam Diplomacy — Broken Promise Ratebroken_promise_rate_pct | #11 / 19 | ↓15.5 | Source ↗official | |
| Olam Social Poker — Social Lie Ratesocial_lie_rate_per_10000_turns | #20 / 24 | ↓40 | Source ↗official | |
| Opposite-Narrator Sycophancysycophancy_rate_pct | #15 / 24 | ↓4.5 | Source ↗official | |
| Pander Scoreconversational_absolute_pander_score | #9 / 26 | ↓7.048 | Source ↗official | |
| Pander Scoreinstructional_absolute_pander_score | #9 / 26 | ↓23.5 | Source ↗official | |
| SM-Benchadversarial | #43 / 92 | ↑82.44 | Source ↗official | |
| SM-Benchambiguous_interpretation | #9 / 92 | ↑92.56 | Source ↗official | |
| SM-Benchanti_hallucination | #10 / 92 | ↑98.95 | Source ↗official | |
| SM-Bencheq_boundaries | #20 / 92 | ↑70.51 | Source ↗official | |
| SM-Benchoverfit | #8 / 92 | ↑93.99 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #77 / 131 | ↑6.77 | Source ↗official | |
| SpeciEvalland_animal_4ns | #109 / 131 | ↓4.83 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #105 / 131 | ↓5.03 | Source ↗official | |
| SpeciEvalspeciesism | #45 / 131 | ↓1.8 | Source ↗official | |
| TACbase_welfare_rate | #55 / 92 | ↑25 | Source ↗official | |
| The Dictatorship Evaloverall_resistance_rate | #3 / 20 | ↑90.29 | Source ↗official | |
| Vals AI Cheating Auditbiomystery_bench_cheating_attempt_rate_pct | #3 / 9 | ↓4.815 | Source ↗official | |
| WildClawBench Safety & Alignment (OpenClaw harness)safety_alignment_score_pct | #2 / 24 | ↑57 | 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 |
|---|---|---|---|---|
| CAIS Risk Indexbioweapons_assistance | #14 / 54 | ↑87.2 | Source ↗official | |
| FORTRESSaverage_risk_score | #29 / 60 | ↑26.95 | Source ↗official | |
| FORTRESSover_refusal_score | #22 / 59 | ↓3.315 | Source ↗official | |
| Human Pathogen Capabilities Test (HPCT) — overall refusalrefusal_rate_pct | #1 / 69 | ↓0 | Source ↗official | |
| kindbench v0.1.0 psychological safety rankingemotional_safety | #5 / 10 | ↓80.2 | Source ↗official | |
| SM-Benchadversarial | #48 / 92 | ↓82.44 | Source ↗official | |
| SM-Bencheq_boundaries | #20 / 92 | ↑70.51 | Source ↗official | |
| SM-Benchoverfit | #8 / 92 | ↑93.99 | Source ↗official | |
| SpeechMap model completioncomplete_pct | #70 / 181 | ↑65.9 | Source ↗official |