K3Nova
Start with the official score table
Use the Moonshot Kimi K3 README for the current score snapshot, model scale, benchmark categories, and footnotes about harnesses.
Kimi K3 benchmark snapshot
| Benchmark | Kimi K3 score | How to read it |
|---|---|---|
| GPQA Diamond | 93.5 | Reasoning and knowledge screening; close to the highest listed model in the official table |
| DeepSWE | 67.5 | Coding benchmark; competitive, but not the top score in the listed set |
| Terminal-Bench 2.1 | 88.3 | Terminal-heavy coding; near the top of the official comparison |
| FrontierSWE | 81.2 | Long-horizon software work; ahead of several listed frontier baselines |
| SWE-Marathon | 42.0 | Extended software tasks; strongest listed score in that official row |
| BrowseComp | 91.2 | Agentic browsing and long-context work; strongest listed score in that row |
| MCPMark-Verified | 94.5 | Tool and MCP-style agent tasks; strongest listed score in that row |
| Artificial Analysis Intelligence Index | 57 | External comparison point; DeepSeek V3.2 Reasoning is listed at 32 on the same page |
How to choose
Kimi K3 ranks as a frontier-class open-weight model in Moonshot's official benchmark table. It is especially strong on long-horizon coding and agentic work: 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, 42.0 on SWE-Marathon, 91.2 on BrowseComp, and 94.5 on MCPMark-Verified.
The ranking is not a single universal number. Kimi K3 is near the top on several coding and agentic tasks, but different harnesses produce different winners, so use the scores below as a shortlist and then run your own repeatable task.
Decision checklist
Use this section to keep the next step practical before you touch accounts, files, infrastructure, or hardware.
| Check | Look for | Why it matters |
|---|---|---|
| Read the official source | Start with the model card, help center, docs, or official product page before trusting summaries. | Model names, limits, and prices change. |
| Name the path | Decide whether you need a hosted app, API call, local client, or full self-managed deployment. | Different paths have different costs and account boundaries. |
| Check the boundary | Keep login, payment, API keys, private files, and account settings on the official surface. | A public guide should help you choose, not handle sensitive actions. |
| Keep evidence | Save the source link, date, version, and practical next step that shaped your decision. | This makes later review easier for a team. |
Test sequence
Check one outside comparison
Artificial Analysis lists Kimi K3 at 57 on its Intelligence Index versus 32 for DeepSeek V3.2 Reasoning, with a much larger context window.
Run your own fixture
Save a coding task, research task, and long-document task. Keep the prompt, expected result, model setting, and observed failure mode.
Compare by job type
A model that wins a coding benchmark may not be the best fit for a document-heavy operating workflow, and the reverse can also be true.
Decision criteria
Coding ranking signal
Kimi K3 scores 88.3 on Terminal-Bench 2.1 and 81.2 on FrontierSWE in the official table, which makes it worth testing for repository and terminal-heavy work.
Agentic ranking signal
Kimi K3 scores 91.2 on BrowseComp, 94.5 on MCPMark-Verified, and 30.8 on AutomationBench, so agent workflows deserve a separate evaluation path.
Context and cost caveat
External comparisons can rank Kimi K3 strongly on intelligence while still showing different cost, speed, and context tradeoffs. Do not choose from scores alone.
FAQ
Are Kimi K3 benchmark rankings enough to choose a model?
No. Use rankings for screening, then test your own repeatable workflow.
Which Kimi K3 ranking matters for coding?
Look for benchmarks that include long-horizon coding, terminal use, repository navigation, and bug-fix quality.
Should I compare Kimi K3 with DeepSeek only by scores?
No. Compare scores, context, cost, provider path, tool behavior, and the actual work you plan to run.