benchmarks
Kimi K2.5 is Moonshot's multimodal model for text generation, image and video understanding, and tool-assisted coding. Standalone model capabilities should be distinguished from tool-enabled execution and Agent Swarm orchestration in the Kimi product.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 96.1%Jan 2026 | View |
| FrontierMath | 27.9%3 Feb 2026 | View |
| GPQA DiamondPhD-level science questions | 87.6%Jan 2026 | View |
| Humanity's Last Exam — without tools | 30.1%Jan 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 92.2%2 Feb 2026 | View |
| SWE-bench VerifiedResolving real GitHub issues | 76.8%Jan 2026 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 50.8%Jan 2026 | View |
Weighted mean of 7 verified benchmark results: AIME 96.1%, FrontierMath 27.9%, GPQA Diamond 87.6%, Humanity's Last Exam — without tools 30.1%, OTIS Mock AIME 2024-25 92.2%, SWE-bench Verified 76.8%, Terminal-Bench 50.8%.
Capabilities
- Text generation and reasoning over long-context inputs.
- Native image and video understanding.
- Coding from screenshots and visual specifications.
- Tool-assisted software engineering and terminal workflows.
- Open-weight deployment for operators with suitable infrastructure.
- Parallel task execution through Agent Swarm orchestration in the Kimi product.
What improved
- Adds native image and video understanding to the text-focused K2 line.
- Extends coding workflows to screenshots, visual specifications, and interface reconstruction.
- Introduces Agent Swarm orchestration for parallel task execution in the Kimi product.
Known flaws & limitations
- The approximately one-trillion-parameter mixture-of-experts architecture requires substantial infrastructure for self-hosting.
- Documented output is text; image and video understanding does not imply native image or video generation.
- Native audio input and output are not documented model modalities.
- Tool-assisted and Agent Swarm results require additional execution infrastructure and do not describe standalone-model performance.
Risks, safety & security issues
- A reported Kimi K3 research demonstration produced Redis remote-code-execution proofs of concept; it does not establish a K2.5 vulnerability.
- Tool-enabled deployments require restricted credentials and execution isolation to limit the consequences of unsafe generated actions.
- Open-weight deployments place access control, abuse monitoring, and enforcement under the operator's control.
Good fit for
- Software engineering with supervised tool execution.
- Interface implementation from screenshots and visual specifications.
- Analysis of documents containing text, images, or video.
- Long-context research and reasoning workflows.
Poor fit for
- Applications requiring native image, video, or audio generation.
- Resource-constrained local deployment.
- Unsupervised execution with unrestricted production credentials.
- Deployments expecting Agent Swarm functionality from model weights alone.
References
Profile last reviewed 10/23/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.