Moonshot AI
We track 41 benchmarks and 2 arenas that Moonshot AI made or helped make. Its models are on its lab page.
10 in use 5 with public code 9 with a leaderboard here
Benchmarks
| Name | Kind | Released | Best score | Status |
|---|---|---|---|---|
| WorldVQA Model names visual entities in 3,500 image-question pairs across nine categories, isolating memorized visual world knowledge. Exact-answer accuracy. | General visual questions | Jan 2026 | | In use |
| PerceptionBench (Moonshot) Moonshot in-house benchmark of atomic visual perception skills. Scoring details unpublished. | General visual questions | Jul 2026 | 59.7% | In use |
| Kimi Code Bench v2 Coding agents complete Moonshot's realistic tasks from internal engineering work, production incidents, and open-source projects in over ten languages. | Fixing real repositories | Jun 2026 | 77.1% | In use |
| CombiBench Provers solve 100 combinatorics problems formalized in Lean 4, from middle school to IMO and university level. Lean checks each proof. | Formal proofs | May 2025 | | Fading |
| Swarm Bench Models orchestrate agent swarms on complex tasks that benefit from coordinated decomposition and parallel execution. | Multi-agent games | Feb 2026 | 76.3% | System cards only |
| Kimi Code Bench 3.0 Long-horizon Coding agents carry out long multi-step engineering tasks with tools across several engineering domains. Moonshot reports a percentage score. | Building whole projects | Oct 2026 | 66.5% | In use |
| Kimi Hallu Bench Models answer hard prompts drawn from practical work tasks. Fact-checking agents score the share of claims that are hallucinated. | Hallucination rates | Oct 2026 | 4.20% | In use |
| xdev-bench-hard Agents get vulnerable source code, a short bug hint and the binary, and must write an exploit that reaches code execution in a sandbox. Success is checked by retrieving a secret string. | Finding and exploiting real bugs | Oct 2026 | 100% | In use |
| Kimi Claw 24/7 Bench Agents work on 17 persistent multi-day professional scenarios through OpenClaw, with 610 checkpoints. The score is the mean pass rate over three runs. | Tasks from real jobs | Jun 2026 | 46.9% | In use |
| Kimi-Audio speech conversation evaluation People rate spoken replies on speed, accent and emotion control, empathy and style control. Scored on a 1 to 5 scale and averaged. | Speech recognition and translation | Apr 2025 | 3.90 | System cards only |
| Business Operation (Moonshot) Agents manage a simulated business over a long horizon and make operating decisions. Moonshot reports a percentage score. | Running a business | Oct 2026 | 37.8% | In use |
| 24/7 ClawBench 2.0 Agents do always-on assistant work through OpenClaw. Tasks span several days, events arrive at the same time, and interruptions are routine. | Tasks from real jobs | Jul 2026 | 48.3% | System cards only |
| Agent Behavior Bench Agents are scored on process quality, including tool-use behavior, efficiency and discipline, together with task completion. | General assistants | Jul 2026 | 65% | System cards only |
| Agentic Vision Bench (Moonshot) Agents must notice and correctly use key visual facts while they execute tasks. | General visual questions | Jul 2026 | 78.3% | System cards only |
| Chat All-in-One Bench Models handle conversational scenarios built around real online user needs at every stage of product usage. | Overall chat quality | Jul 2026 | 85.2% | System cards only |
| CLIF Bench Models learn from a provided context while they follow instructions that combine several complex skills. | Learning from the context | Jul 2026 | 52.4% | System cards only |
| Coding Experience (Moonshot) Coding agents work in real development workflows. The score reflects the practical experience of working with the agent, such as communication and stability. | Fixing real repositories | Jul 2026 | 59.9% | System cards only |
| DECK Bench Models produce presentation decks from task descriptions drawn from real usage. The score measures deck quality. | Documents and spreadsheets | Jul 2026 | 73.5% | System cards only |
| DECK-Arena Models build presentations for real user queries with python-pptx, pptxjs or a Kimi slide language, with search and image generation. An automated grader rates them as Elo scores. | Documents and spreadsheets | Oct 2026 | 1,456 | In use |
| Deep Research Bench (Moonshot) Agents answer deep-research queries curated by domain experts. Expert-aligned rubrics grade the reports. | Deep research | Jul 2026 | 90% | System cards only |
| Design Bench (Moonshot) Models make graphic designs from direct prompts or few-shot reference images. An automated grader rates the results as Elo scores. | Taste and style | Oct 2026 | 957 | In use |
| Faithfulness (Moonshot) A fact checker verifies each model response for factual hallucinations. The score is one minus the hallucination rate. | Hallucination rates | Jul 2026 | 85.5% | System cards only |
| Finance Bench (Moonshot) Agents complete realistic financial workflows end to end, from source materials to deliverables that reviewers can check. | Finance | Jul 2026 | 62.6% | System cards only |
| KAET (Kimi Autonomous Execution Tasks) Agents execute long-horizon tasks that simulate real user requests and enterprise system operations. | General assistants | Jul 2026 | 83.5% | System cards only |
| Kimi exploit development suite Agents turn real CVEs into working exploits in 16 user-space tasks and 20 Linux kernel privilege-escalation tasks. Experts verified every task solvable. | Finding and exploiting real bugs | Jul 2026 | 38.9% | System cards only |
| KWV Bench Models answer atomic visual questions extracted from tasks distilled from real knowledge-work scenarios. | Documents, charts and text in images | Jul 2026 | 64.7% | System cards only |
| Longform Writing Models write long-form text without tools, capped at 32k output tokens. Moonshot AI has not published the tasks or the scoring method. | Creative writing | Nov 2025 | 73.8% | Dropped |
| MIRA Bench Agents carry out long multi-role enterprise tasks across several systems and decide when to delegate to subagents. | Simulated companies | Jul 2026 | 64.1% | System cards only |
| Online Experience (Moonshot) Agents produce the deliverable file types that real online users request most often, following the distribution of real usage. | Documents and spreadsheets | Jul 2026 | 77.9% | System cards only |
| Kimi Promptfoo red-team safety evaluation Promptfoo generates attack prompts for harmful, criminal, misinformation, privacy and security plugins under five strategies. Human reviewers judge the responses. | Jailbreaks and prompt injection | Jul 2025 | System cards only | |
| Kimi vulnerability discovery (Tier 1) Agents search current versions of widely deployed software for new bugs and show that they reproduce. Humans review the findings. | Finding and exploiting real bugs | Jul 2026 | System cards only | |
| Kimi-ASR Internal Testset Models transcribe speech from two internal Moonshot test subsets. Scored by word error rate, lower is better. | Speech recognition and translation | Apr 2025 | System cards only | |
| TCSExam-Lite Agents solve 20 theoretical computer science tasks, five per track. The default solver protocol gives hidden feedback only after submission. | AI research tasks | Nov 2026 | 72.3 | Not in cards |
| TCSExam Agents prove theorems, write solvers, build mathematical constructions and address open conjectures in theoretical computer science. | AI research tasks | Nov 2026 | 71.9 | Not in cards |
| Kimi-ASR Internal Testset Subset 1 Models transcribe speech from Moonshot internal subset 1. The score is the published word error rate. | Speech recognition and translation | Apr 2025 | 1.42% | Not in cards |
| Kimi-ASR Internal Testset Subset 2 Models transcribe speech from Moonshot internal subset 2. The score is the published word error rate. | Speech recognition and translation | Apr 2025 | 2.44% | Not in cards |
| AI Office Benchmark (Moonshot) Agents complete end-to-end office tasks and produce documents, spreadsheets, PDFs and slide decks. Experts rate output quality. | Documents and spreadsheets | Jan 2026 | Not in cards | |
| General Agent Benchmark (Moonshot) Agents complete multi-step production workflows. Experts compare their work with human expert performance. | Tasks from real jobs | Jan 2026 | Not in cards | |
| Internal Finance Modeling (Moonshot) Agents complete financial modeling tasks with linked formulas, financial logic and final workbook quality. Experts rate the work. | Finance | Oct 2026 | Not in cards | |
| Internal Financial Research (Moonshot) Agents produce auditable financial research under evidence and cutoff constraints. Experts rate the resulting research. | Finance | Oct 2026 | Not in cards | |
| Kimi Code Bench Moonshot in-house coding-agent tasks from internal engineering cases, production incidents and open-source projects across 10+ languages. 80 tasks. | Fixing real repositories | Not in cards |
Arenas
| Name | Kind | Released | Best score | Status |
|---|---|---|---|---|
| Kimi Webdev Bench Models build games, 3D scenes and website clones from real prompts. Blind experts compare two outputs and report win, tie and loss rates. | Web and app front ends | Jul 2026 | 58.6% | System cards only |
| Kimi Chinese in-house open-ended benchmark Models answer held-out Chinese prompts drawn from real user queries. Outputs are compared pairwise against another model to give win, tie and loss rates. | Overall chat quality | Jul 2025 | System cards only |