HumanEval
View DetailsA benchmark of 164 hand-written programming problems used to evaluate functional code generation.
LanguageEnglishDifficultyEasyCoding and Software EngineeringInstitutionOpenAI
Benchmark Library
This page aggregates mainstream LLM evaluation benchmarks including AIME 2025, SWE Bench Verified, MMLU, GSM8K, HumanEval, and more. We provide a comprehensive reference platform for researchers and developers to understand model performance across various evaluation datasets.
All Benchmarks
117
All categories
19
All languages
5
Detailed evaluation results on benchmark leaderboards:View Benchmark Leaderboards
117 benchmarks matched. Recommended flow: review leaderboards, then compare models.
A benchmark of 164 hand-written programming problems used to evaluate functional code generation.
Mostly Basic Python Problems
A benchmark of 974 short Python programming problems used to evaluate code generation.
Software Engineering Bench
A software-engineering benchmark built from real Python repositories and GitHub issues to test whether models can resolve practical coding tasks.
Software Engineering Bench - Verified
A human-validated subset of SWE-bench curated to provide more accurate and representative software-engineering tasks.
Individual Contributor SWE-Lancer(Diamond)
A benchmark that evaluates whether models can complete realistic software-engineering tasks performed by individual contributors.
SWE Manager-Lancer(Diamond)
A benchmark that evaluates model performance on realistic software-engineering management tasks.
A continuously updated benchmark that uses difficult problems from competitive-programming platforms to evaluate coding ability while reducing data contamination.
A benchmark that uses competitive-programming problems to evaluate code generation and algorithmic reasoning.
Software Engineering Bench Pro - Public Dataset
A public benchmark for evaluating whether models can solve realistic, complex software-engineering tasks.
Software Engineering Bench Pro - Commercial Dataset
A commercial benchmark dataset for evaluating whether models can solve realistic, complex software-engineering tasks.
A multilingual software-engineering benchmark covering realistic repository tasks across nine programming languages.
DeepSWE: Long-Horizon Software Engineering Benchmark
A long-horizon software engineering benchmark from Datacurve with 113 original tasks across 91 active open-source repositories and five programming languages. It emphasizes multi-file changes, behavioral correctness, and reduced contamination risk.
CodeClash: Benchmarking Goal-Oriented Software Engineering
An open benchmark for goal-oriented software engineering. Models maintain and improve codebases through multi-round tournaments, with win rate and Elo measuring strategic planning, log analysis, code evolution, and long-term maintenance.
Artificial Analysis Coding Agent Index
Artificial Analysis 面向编程智能体（agentic coding）场景的第三方综合评测指数。
SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?
SWE-Marathon 是面向超长周期软件工程智能体的评测基准，包含 20 个跨软件工程及相关技术领域的多小时任务，考察持续规划、长上下文理解、记忆、自我验证和复杂环境执行能力，并使用多层验证与对抗审查降低捷径解和奖励劫持的影响。
Program Bench 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
FrontierSWE 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
PostTrain Bench 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
MLS Bench Lite
MLS Bench 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
Kimi Code Bench 2.0（Kimi 内部评测） 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
SciCode 是面向科研代码生成与问题求解的评测基准；Kimi K3 模型卡引用 Artificial Analysis 截至 2026-07-23 的结果。
NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents
NL2Repo-Bench evaluates whether coding agents can build a complete, installable Python repository from a natural-language specification and an empty workspace. The 104 tasks are graded against upstream pytest suites.
Massive Multitask Language Understanding
A multiple-choice benchmark spanning 57 subjects that evaluates the knowledge and reasoning capabilities of large language models.
Massive Multitask Language Understanding Professional Level
A more challenging, professional-level variant of MMLU designed to evaluate model understanding and reasoning in specialized domains.
Beyond the Imitation Game Benchmark
A broad benchmark with more than 200 tasks covering reasoning, language understanding, knowledge, and other model capabilities.
Chinese Evaluation Benchmark
A Chinese multiple-choice benchmark spanning humanities, social sciences, and STEM subjects to evaluate knowledge and reasoning in Chinese.
Big Bench Hard
A difficult subset of BIG-bench containing especially challenging tasks that test the limits of model capabilities.
An abstract-reasoning benchmark created by the founder of Keras to evaluate progress toward general intelligence.
Humanity's Last Exam
An exceptionally difficult multidisciplinary benchmark with questions at graduate level and beyond.
The highest-quality subset of GPQA, designed to evaluate expert-level knowledge and reasoning across difficult scientific domains.
A Graduate-Level Google-Proof Q&A Benchmark
A graduate-level, Google-proof question-answering benchmark designed to evaluate expert knowledge and rigorous reasoning.
A highly challenging abstract-reasoning benchmark designed to measure general problem-solving ability.
A contamination-resistant benchmark designed to provide frequently updated and objective evaluation of large language models.
A general abstract-reasoning benchmark designed to evaluate adaptive problem solving and machine intelligence.
Artificial Analysis Intelligence Index
Artificial Analysis 第三方综合智能评测指数，覆盖推理、知识、编程等多维度任务，按厂商公开 API 独立测试。
Vals AI Index
Vals AI 第三方综合评测指数，覆盖法律、金融、医疗等真实行业任务的准确率测试。
CritPt 是 Kimi K3 开放权重模型卡采用的推理与知识基准；K3 成绩由模型卡引用 Artificial Analysis 截至 2026-07-23 的结果。
A benchmark for evaluating the agentic ability of models to complete practical tasks by using tools in a terminal environment.
Terminal Bench Hard 2.0
An upgraded terminal-environment benchmark for evaluating the tool-use and task-completion abilities of language-model agents.
A benchmark that evaluates whether language models can use tools to complete complex tasks in realistic environments.
OS World Verified
A verified successor to OSWorld for evaluating the ability of language-model agents to operate computers in realistic environments.
A benchmark developed by Stanford University and the Laude Institute to evaluate AI agents in realistic terminal environments.
MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers
A large-scale Scale AI benchmark for tool use with real MCP servers. It covers 36 servers, 220 tools, and 1,000 natural-language tasks that test tool discovery, cross-server orchestration, multi-step calls, and answer synthesis.
Vals AI 网络安全专项评测子榜单，测试模型在漏洞分析、渗透测试等网络安全任务中的表现。
OSWorld 2.0 评测计算机操作智能体在真实桌面环境中完成任务的成功率，数值经二手来源 MarkTechPost 转引 OpenAI system card 数据，与 OSWorld-Verified 为不同版本/口径，独立建档。
Toolathlon-Verified 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
AutomationBench
Automation Bench 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering
MLE-Bench 用 75 个 Kaggle 竞赛任务评估 AI 智能体执行端到端机器学习工程工作的能力，覆盖数据准备、模型训练、实验运行与结果改进。Gemini 3.6 Flash 官方模型卡使用该基准报告模型对比成绩。
MCPMark-Verified 用于评估智能体通过 MCP 工具完成真实任务的能力；Kimi K3 成绩来自官方开放权重模型卡。
SaaS-Bench 用于评估智能体操作软件即服务应用并完成跨步骤任务的能力；Kimi K3 成绩来自官方开放权重模型卡。
CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale
CyberGym is a large-scale cybersecurity evaluation framework for assessing AI agents on real-world vulnerability analysis tasks in reproducible execution environments.
Grade School Math 8K
A benchmark of 8,500 grade-school mathematics word problems used to evaluate mathematical reasoning.
Mathematics Problem Solving
A challenging mathematics dataset covering algebra, calculus, geometry, probability, and other fields.
A representative set of 500 problems selected from the MATH dataset to evaluate advanced mathematical reasoning.
The 2024 American Invitational Mathematics Examination, consisting of 15 challenging problems completed in a three-hour session.
Problems from the 2025 American Invitational Mathematics Examination used to evaluate advanced mathematical reasoning.
International Olympiad in Informatics 2024
A 2024 International Olympiad in Informatics benchmark used to evaluate advanced algorithmic problem solving and code generation.
International Olympiad in Informatics 2025
A 2025 International Olympiad in Informatics benchmark used to evaluate advanced algorithmic problem solving and code generation.
FrontierMath (Tier 1-3)
A benchmark of hundreds of novel and exceptionally difficult problems written by mathematicians to measure the frontier of advanced mathematical reasoning.
The most difficult FrontierMath tier, featuring research-level mathematical problems intended to test frontier reasoning systems.
International Mathematics Olympiad-ProofBench
A benchmark for evaluating whether large language models can generate complete, logically rigorous proofs for Olympiad-level mathematics problems.
International Mathematics Olympiad-ProofBench Advanced
An advanced benchmark for evaluating complete and logically rigorous proofs generated for difficult Olympiad-level mathematics problems.
International Mathematical Olympiad-AnswerBench
A benchmark that tests answers to difficult International Mathematical Olympiad-level mathematics problems.
American Invitational Mathematics Examination 2026
A benchmark based on the 2026 American Invitational Mathematics Examination for evaluating advanced mathematical reasoning.
Massive Multi-discipline Multimodal Understanding and Reasoning
A college-level multimodal benchmark that evaluates cross-disciplinary reasoning and the application of expert knowledge to text-and-image problems.
Simple Visual Question Answering
A visual question-answering benchmark focused on the factual accuracy of short natural-language answers from multimodal models.
Document Visual Question Answering
A visual question-answering benchmark built around document images and document-understanding tasks.
MMMU-Pro 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
CharXiv (RQ)
CharXiv RQ 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
MathVision 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
BabyVision 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
ZeroBench_main
ZeroBench Main 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
WorldVQA ForceAnswer 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
OmniDocBench 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
PerceptionBench（Kimi 内部评测） 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
Video-MME with Subtitles
Video-MME（带字幕）评估模型对视频内容与字幕信息的综合理解能力；Kimi K3 成绩来自官方开放权重模型卡。
MMVU 是多模态视频理解评测；Kimi K3 成绩来自官方开放权重模型卡。
Gross Domestic Product Value - Artificial Analysis
A benchmark that measures performance on knowledge-work tasks drawn from industries that contribute heavily to United States GDP.
GDPval-AA v2 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
Artificial Analysis Briefcase
AA-Briefcase 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
OfficeQA Pro
Office QA Pro 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
SpreadsheetBench 2 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
DECK-Bench（Kimi 内部评测） 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
ResearchRubrics 是 Kimi K3 开放权重模型卡报告的智能体研究能力评测，具体测试设置以官方模型卡和技术报告为准。
Harvey Lab - Artificial Analysis
Harvey Lab-AA 是 Artificial Analysis 的专业法律知识工作评测；Kimi K3 模型卡报告 criterion pass rate。
Corporate Finance v2
CorpFin v2 是 Vals AI 的企业财务知识工作评测；Kimi K3 成绩由官方模型卡引用。
Finance Agent v2 是 Vals AI 的金融智能体知识工作评测；Kimi K3 成绩由官方模型卡引用。
Legal Research Bench 是 Vals AI 的法律研究知识工作评测；Kimi K3 成绩由官方模型卡引用。
Aider-Polyglot Benchmark
A multilingual benchmark that evaluates whether models can follow instructions to make practical, correct edits to existing code.
τ²-Bench - Overall
An open benchmark for evaluating task execution and communication by language-model agents in dual-control human-agent environments.
A telecom-domain benchmark that evaluates task execution by language-model agents in human-agent collaboration scenarios.
An Artificial Analysis benchmark variant focused on difficult tool-use tasks in terminal environments.
GAIA: A Benchmark for General AI Assistants
A benchmark for general AI assistants introduced by researchers from Meta, Hugging Face, and partner institutions. Its 466 real-world questions require reasoning, multimodal understanding, web browsing, and tool use.
Agents' Last Exam evaluates AI agents on long-horizon, economically valuable professional workflows with verifiable outcomes across a broad set of industries.
JobBench
Job Bench 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
APEX-Agents 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
τ³-Benchmark - Banking
τ³-Banking 是 Kimi K3 模型卡采用的银行业务智能体评测，考察多轮交互和工具驱动任务完成能力。
Artificial Analysis - Long Context Reasoning
An Artificial Analysis benchmark that simulates knowledge workers such as analysts, researchers, and lawyers reasoning over large document collections.
LongBench v2: Deeper Understanding and Reasoning on Realistic Long-Context Multitasks
A realistic long-context understanding and reasoning benchmark with 503 multiple-choice questions spanning contexts from 8,000 to 2 million words. Tasks include single- and multi-document QA, long conversations, code repositories, and structured data.
Google DeepMind MRCR v2 8-needle - 128K Average
GDM-MRCR v2（8-needle）用于评估模型从长上下文中检索并整合多个关键信息的能力。本目录项对应 Google DeepMind 模型卡中的 128K average 测试条件，与 1M pointwise 口径分开记录。
Google DeepMind MRCR v2 8-needle - 1M Pointwise
GDM-MRCR v2（8-needle）用于评估模型从长上下文中检索并整合多个关键信息的能力。本目录项对应 Google DeepMind 模型卡中的 1M pointwise 测试条件，与 128K average 口径分开记录。
Hella Swag
A benchmark of roughly 70,000 multiple-choice questions designed to evaluate commonsense reasoning.
AI2 Reasoning Challenge
A benchmark of 7,787 multiple-choice science questions used to evaluate scientific and commonsense reasoning.
A benchmark designed to evaluate the commonsense reasoning capabilities of large language models.
A benchmark that evaluates the ability of AI agents to locate difficult-to-find information through web research.
DeepSearchQA 是 Kimi K3 官方技术博客完整评测表采用的基准。本条目用于承载官方发布成绩；Kimi K3 的全局测试设置为 reasoning_effort=max、temperature=1.0、top_p=1.0，具体 harness、工具使用和重复运行口径以官方脚注为准。
A benchmark that evaluates the ability of a language model to serve as the control center for OpenClaw.
A benchmark that evaluates language models across five common categories of OpenClaw tasks.
Massive Text Embedding Benchmark
A comprehensive benchmark suite for evaluating text-embedding models across retrieval, classification, clustering, and related tasks.
Context Arena
A benchmark that uses multi-turn conversations to evaluate long-context understanding and reasoning.