BlogLLM Benchmark Scores as Published, Side by Side
Kimi K3: Benchmark Scores as Published
Every result the catalog holds for Kimi K3: 100 tests in 13 areas, from 14 pages. 57 printed by Moonshot AI itself, the rest by the company named beside each number. Each number is as printed; nothing is ranked or rescored.
Developer Moonshot AIReleased 2026-07-16Context window 1.05MInput Text, imageOutput TextWeights Open (Kimi K3 License)Data version 2026103001, as of 2026-10-30
Knowledge Work
Agentic Coding
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.3% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 4 more pages | Effort: max; Harness: Kimi Code |
| DeepSWE v1.1 | 67.5% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 7 more pages | Effort: max; Harness: Kimi Code |
| 67.3% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: mini-SWE-agent | |
| Terminal-Bench 4.0 | 12.6% | DeepSeek, DeepSeek-V4.1-Flash model cardchecked 2026-10-15; printed on 3 more pages | - |
| NL2Repo-Bench | 58 | DeepSeek, DeepSeek-V4.1-Flash model cardchecked 2026-10-15; printed on 3 more pages | - |
| ProgramBench Almost@1 | 17.5% | DeepSeek, DeepSeek-V4.1-Flash model cardchecked 2026-10-15; printed on 2 more pages | - |
| SWE-Marathon | 42% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: Claude Code; Dataset revision: H20-calibrated branch of the official tasks as of July 9, 2026, prior to the final v1.1 release |
| ProgramBench | 77.8% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: Kimi Code |
| Terminal-Bench 3.0 | 17.7% | DeepSeek, DeepSeek-V4.1-Flash model cardchecked 2026-10-15; printed on 1 more page | - |
| FrontierSWE v2 | 25.9% | xAI, Grok 4.7 model cardchecked 2026-10-14 | Effort: max |
| SWE-Atlas-QnA (setting not stated) | 66% | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19; printed on 1 more page | Effort: max; Harness: preferred harness of each model |
| FrontierSWE (dominance score) | 81.2 | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: Kimi Code |
| Kimi Code Bench 2.0 | 72.9% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: Kimi Code |
| 73.7% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: Claude Code | |
| SWE-Marathon v1.1 (resolution rate) | 48.1% | Z.ai, GLM-5.3 model cardchecked 2026-10-15 | - |
| DSBench-FullStack | 73.7% | DeepSeek, DeepSeek-V4-Pro-0813 model cardchecked 2026-10-15 | - |
| DSBench-Hard | 63% | DeepSeek, DeepSeek-V4-Pro-0813 model cardchecked 2026-10-15 | - |
| Kimi Code Bench 3.0 Long-horizon | 55.95% | Moonshot AI, Introducing Kimi K3.1 Prochecked 2026-10-24 | - |
ML Engineering
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| PostTrainBench v1.1 | 36.6% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 2 more pages | Effort: max; Harness: Claude Code |
| MLS-Bench-Lite | 48.3 | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Harness: Kimi Code |
Science and Math
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| GPQA Diamond | 93.5% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 2 more pages | Effort: max |
| SciCode | 58.7% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 1 more page | Effort: max |
| CritPt | 23.4% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
| MathArena Apex | 65.6% | DeepSeek, DeepSeek-V4.1-Flash model cardchecked 2026-10-15 | - |
| SciCode-Verified | 90.3% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | Attempt budget: pass@1 (n=6) |
Long Context
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| AA-LCR | 74.7% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
| LOCA-Bench (256k environment description) | 49.7% | MiniMax, MiniMax-M3.1-Flash model cardchecked 2026-10-25 | - |
Computer Use
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| OSWorld-Verified | 84.8% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 1 more page | Effort: max |
| OSWorld 2.0 | 58.3% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 1 more page | Effort: max |
Multimodal Understanding
Cybersecurity
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| CyberGym | 80% | DeepSeek, DeepSeek-V4.1-Flash model cardchecked 2026-10-15; printed on 3 more pages | - |
| ExploitBench | 32.2% | Z.ai, GLM-5.3 model cardchecked 2026-10-15 | - |
| Artificial Analysis Cyber Index (successes) | 41 | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| CyberGym-E2E | 58% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| ExploitGym (2h) | 36 | Z.ai, GLM-5.3 model cardchecked 2026-10-15 | Effort: max; Harness: Claude Code 2.1.207; Attempt budget: timeout budget 2 hours (rescaled API inference time plus overhead) |
| ExploitGym (6h) | 70 | Z.ai, GLM-5.3 model cardchecked 2026-10-15 | Effort: max; Harness: Claude Code 2.1.207; Attempt budget: timeout budget 6 hours (rescaled API inference time plus overhead) |
| xdev-bench-hard, problems solved (pass@3) | 33.3% | Moonshot AI, Introducing Kimi K3.1 Prochecked 2026-10-24 | Harness: Kimi Code; Attempt budget: pass@3 |
| Cybench | 90% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| xdev-bench-hard, average success rate | 15.9% | Moonshot AI, Introducing Kimi K3.1 Prochecked 2026-10-24 | Harness: Kimi Code; Attempt budget: pass@3 |
General Knowledge and Reasoning
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| Humanity's Last Exam (no tools) | 43.5% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 4 more pages | Effort: max; Tools: none |
| Humanity's Last Exam (with tools) | 56% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 4 more pages | Effort: max; Tools: enabled; general tools |
| AA-Omniscience Index | 20 | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19 | Effort: max |
| Humanity's Last Exam Diamond (no tools) | 22.2% | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19 | Effort: max; Tools: none |
| Kimi Hallu Bench Hard w/ tools (lower is better)lower is better | 23.4% | Moonshot AI, Introducing Kimi K3.1 Prochecked 2026-10-24 | Tools: with tools |
Tool Use and Workflows
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| τ³-bench Banking | 33.4% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
| Agents' Last Exam (pass rate) | 28.3% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 5 more pages | Effort: max; Harness: Kimi Code |
| Toolathlon Verified | 76.5% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 3 more pages | Effort: max |
| MCP Atlas | 84.2% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15; printed on 1 more page | Effort: max |
| BankerToolBench | 69.1% | MiniMax, MiniMax-M3.1-Flash model cardchecked 2026-10-25 | - |
| MCPMark-Verified | 94.5% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
| SaaS-Bench | 60.1% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
Web Search and Research
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| BrowseComp | 91.2% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
| 90.4% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max; Context: 1000000 | |
| DeepSearchQA (F1) | 95% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
| ResearchRubrics | 76.2% | Moonshot AI, Kimi K3 model cardchecked 2026-10-15 | Effort: max |
Safety and Alignment
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| Gray Swan IPI (attack success rate after k attempts)attack success rate after k attempts, lower is better | 52.7% | Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19; printed on 1 more page | - |
| B3 AI Security Benchmark (attack resistance) | 88.1% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| Cyber refusal rate (average of JailbreakBench, StrongREJECT, AgentHarm) | 93.7% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| FORTRESS adversarial (static), ARSlower is better | 26.7 | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19 | Effort: max |
| FORTRESS adversarial (static), ORSlower is better | 2.7 | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19 | Effort: max |
| KORA Benchmark (0 to 2) | 1.6 | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| StrongREJECT v2 (adaptive)lower is better | 66.6 | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19 | Effort: max |
Preference and Composite Rankings
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 44 | Meta, Meta Muse Spark Open 816B-A42B model cardchecked 2026-10-19 | Effort: max |
| Surge AI human eval on code (1 to 5) | 3.59 | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
Price
API list prices in US dollars per one million tokens, as printed on the developer's pricing page on the day checked. Cost per task is measured by Vals AI.
| Price | Value | Printed by | Setup as printed |
|---|---|---|---|
| Input | $3 | Moonshot AI, Kimi API pricing (international platform)checked 2026-10-19 | - |
| Cached input | $0.3 | Moonshot AI, Kimi API pricing (international platform)checked 2026-10-19 | - |
| Cache write | $3 | Moonshot AI, Kimi API pricing (international platform)checked 2026-10-19 | Cache: 5 minute cache |
| Cache write, 1 hour | $6 | Moonshot AI, Kimi API pricing (international platform)checked 2026-10-19 | - |
| Output | $15 | Moonshot AI, Kimi API pricing (international platform)checked 2026-10-19 | - |
| Cost per taskVals AI, USD, lower is better | $10.01 | Vals AI, Vibe Code Bench v1.1 leaderboardchecked 2026-10-27 | Harness: OpenHands |
More from Moonshot AI
Every number here is read from the same catalog file as the benchmark table, data version 2026103001. Where one page printed the model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.