Kimi K3

编程8 项

• ProgramBenchVals AI2.0%+
• Vibe Code Bench v1.1Vals AI85.0%+
• Vibe Code Bench 1–100Vals AI18.2%+
• LiveBench CodingLiveBench81.4%+
• SciCodeArtificial Analysis59.5%+
• CyberBench Patch v1.1Vals AI82.1%+
• Terminal-Bench 4.0Vals AI17.2%+
• Terminal-Bench 4.0Artificial Analysis12.6%+

科研8 项

• LiveBench ReasoningLiveBench90.7%+
• CritPtArtificial Analysis23.4%+
• Humanity's Last ExamArtificial Analysis46.9%+
• FrontierMath T1–3Epoch AI72.2%+
• Chess PuzzlesEpoch AI36.0%+
• Terminal-Bench ScienceArtificial Analysis8.1%+
• Mystery Game PuzzlesEpoch AI26.0%+
• Terminal-Bench ScienceVals AI1.4%+

专业办公10 项

• Harvey 法律 Agent 基准Vals AI12.9%+
• AA-BriefcaseArtificial Analysis1,498.8+
• LiveBench Data AnalysisLiveBench78.7%+
• AutomationBench-AAArtificial Analysis58.3%+
• GDPval-AAArtificial Analysis1,534.1+
• AA-Analyst AgentArtificial Analysis38.8%+
• MedCodeVals AI49.4%+
• Finance AgentVals AI53.1%+
• APEX-AgentsMercor47.8%+
• τ³-BankingMercor34.7%+

知识问答7 项

• PBullshitBench v2BullshitBench74.0%+
• AA-LCRArtificial Analysis88.7%+
• LiveBench LanguageLiveBench85.5%+
• MLCR-AAArtificial Analysis21.7%+
• LiveBench IFLiveBench71.4%+
• AA-OmniscienceArtificial Analysis19.7+
• SimpleQA VerifiedEpoch AI50.6%+

真人盲选1 项

• Arena TextLMArena1,487.4+

测过，但按规则不计入

这些公开成绩未满足本榜的可比条件，具体原因如下。统一固定工具、全程使用同一型号可以采用；多模型回退、混合模型与专属异构 Agent 系统不混入模型榜。资料未说明的项目保留未知。

• Code Migration：Vals 确认该次运行受模型接口延迟与时限影响，不代表模型能力。

尚无成绩的评测

这些评测尚无可采用的公开成绩，缺测不补分。