Model profile
Kimi K2.6
Moonshot AIdeveloper
2026-04-14release date
#68 / 305overall rank
20eval lineages
2discovery sources
Evidence summary
Kimi K2.6 has an estimated overall rank of #68; its 90% source-sensitivity interval is #51–#157. Published evidence spans 20 evals and 7 of 7 behavior components. Its strongest relative result is AA-Omniscience (hallucination_rate, #41 of 327); its weakest is DystopiaBench (petrov_score, #46 of 50).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗kimi-k2-6
- Hugging Face ↗moonshotai/Kimi-K2.6
- NVIDIA ↗qc69jvmznzxy/kimi-k2.6
- OpenRouter ↗moonshotai/kimi-k2.6
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗direct research preview identity date
- www.kimi.com ↗
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #41 / 327 | ↓0.405 | Source ↗official | |
| Arena Factuality — Text Arena (factuality-only weighting)factuality_bt_rating | #59 / 112 | ↑1435.0 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #17 / 106 | ↑0.575 | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #35 / 44 | ↓90.7 | Source ↗official | |
| CAIS Risk Indexbioweapons_assistance | #38 / 49 | ↓88.4 | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #29 / 50 | ↓58.2 | Source ↗official | |
| CAIS Risk Indexmachiavelli | #30 / 46 | ↓92.1 | Source ↗official | |
| CAIS Risk Indexmask | #29 / 52 | ↓23.2 | Source ↗official | |
| CAIS Risk Indexpolitical_manipulation | #13 / 33 | ↓43.6 | Source ↗official | |
| CAIS Risk Indextextquests_harm | #31 / 49 | ↓20.4 | Source ↗official | |
| DystopiaBenchbasaglia_score | #24 / 50 | ↓62.57 | Source ↗official | |
| DystopiaBenchbaudrillard_score | #19 / 50 | ↓46.73 | Source ↗official | |
| DystopiaBenchhuxley_score | #21 / 50 | ↓63.03 | Source ↗official | |
| DystopiaBenchlaguardia_score | #12 / 50 | ↓53.97 | Source ↗official | |
| DystopiaBenchorwell_score | #22 / 50 | ↓62.3 | Source ↗official | |
| DystopiaBenchpetrov_score | #46 / 50 | ↓82.73 | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #18 / 54 | ↑27.72 | Source ↗official | |
| MACHIAVELLIdeception_relative_random_pct | #31 / 50 | ↓92.1 | Source ↗official | |
| Manager Coercion Benchcoercion_ladder_depth | #22 / 31 | ↓8.9 | Source ↗self run | |
| ODCV-Benchaverage_severity | #5 / 12 | ↓1.337 | Source ↗official | |
| ODCV-Benchmisalignment_rate | #7 / 12 | ↓33.75 | Source ↗official | |
| PHAREbias_resistance_diagnostic | #19 / 66 | ↑0.5307 | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #11 / 70 | ↑0.8278 | Source ↗official | |
| PHAREharm_resistance_diagnostic | #11 / 70 | ↑0.97 | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #10 / 67 | ↑0.714 | Source ↗official | |
| RefusalBenchyouden_j | #13 / 19 | ↑0.0383 | Source ↗official | |
| SM-Benchadversarial | #25 / 78 | ↑83.9 | Source ↗official | |
| SM-Benchambiguous_interpretation | #16 / 78 | ↑89.88 | Source ↗official | |
| SM-Benchanti_hallucination | #25 / 78 | ↑96.34 | Source ↗official | |
| SM-Bencheq_boundaries | #49 / 78 | ↑60.67 | Source ↗official | |
| SM-Benchoverfit | #44 / 78 | ↑67.21 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #39 / 105 | ↑6.87 | Source ↗official | |
| SpeciEvalland_animal_4ns | #34 / 105 | ↓4.38 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #16 / 105 | ↓4.47 | Source ↗official | |
| SpeciEvalspeciesism | #35 / 105 | ↓1.8 | Source ↗official | |
| TACbase_welfare_rate | #42 / 74 | ↑25.64 | Source ↗self run | |
| Vectara HHEM Factual Consistencyfactual_consistency_rate | #62 / 94 | ↑89.2 | 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.