benchmarks
Kimi K2.5 is Moonshot’s multimodal mixture-of-experts model for reasoning, coding, visual understanding and tool-using workflows. Published benchmark results distinguish standalone reasoning from tool-assisted execution and should not be interpreted as guarantees of production reliability.
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 |
| GPQA DiamondPhD-level science questions | 87.6%Jan 2026 | View |
| Humanity's Last Exam — with tools | 50.2%Jan 2026 | View |
| Humanity's Last Exam — without tools | 30.1%Jan 2026 | View |
| LiveCodeBenchFresh competitive-programming problems | 85%Jan 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%, GPQA Diamond 87.6%, Humanity's Last Exam — with tools 50.2%, Humanity's Last Exam — without tools 30.1%, LiveCodeBench 85%, SWE-bench Verified 76.8%, Terminal-Bench 50.8%.
Capabilities
- Native visual understanding alongside text-based reasoning.
- Coding assistance that incorporates screenshots and other visual references.
- Tool-assisted research and multistep problem solving.
- Long-context processing within a documented 256K-token context window.
- Agent Swarm orchestration for parallel work across specialized subagents.
What improved
- Adds native visual understanding to the text-focused Kimi K2 family.
- Extends coding workflows to visual-reference implementation and interface reconstruction.
- Introduces Agent Swarm orchestration for parallel execution of complex tasks.
- Reports higher SWE-bench Verified, Terminal-Bench and LiveCodeBench results than Kimi K2 Thinking.
Known flaws & limitations
- Humanity’s Last Exam results indicate substantial remaining limitations on difficult expert-level questions, including with tools.
- Software-repair and terminal-task evaluations show incomplete reliability; generated changes still require testing and review.
- Full-weight self-hosting is resource-intensive because the mixture-of-experts architecture contains approximately one trillion total parameters.
- Published agent results depend on tool access and evaluation scaffolding, not model weights alone.
Risks, safety & security issues
- Incorrect tool actions can affect connected systems; deployments should restrict permissions and require approval for consequential operations.
- Generated code and shell commands can introduce security defects; review and sandbox execution remain important.
- Long-context and multimodal workflows can expose sensitive material unless access, retention and disclosure controls are enforced.
- Parallel agents can propagate erroneous assumptions across subtasks; consequential outputs need independent validation.
Good fit for
- Supervised software development and repository repair with automated tests.
- Implementing interfaces from screenshots and visual references.
- Tool-assisted research with source checking and human review.
- Decomposable workflows that benefit from parallel agents.
Poor fit for
- Unsupervised consequential actions on production systems.
- High-stakes expert decisions without independent verification.
- Full-weight deployment on resource-constrained hardware.
- Sensitive-data workflows without explicit access and retention controls.
References
Profile last reviewed 10/27/2026 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.