benchmarks
Kimi K2 Thinking is Moonshot's open-weight, text-only reasoning model, designed for coding, complex problem solving and tool-assisted workflows with a 256K-token context window. Its published benchmark results include tool-assisted evaluations, which should not be interpreted as unaided model performance.
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 | 99.1%Nov 2025 | View |
| GPQA DiamondPhD-level science questions | 84.5%Nov 2025 | View |
| Humanity's Last Exam — with tools | 44.9%Nov 2025 | View |
| Humanity's Last Exam — without tools | 26.9%Nov 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 83.1%Nov 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 71.3%Nov 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 35.7%6 Nov 2025 | View |
| FrontierMath | 21.4%5 Dec 2025 | View |
Weighted mean of 8 verified benchmark results: AIME 99.1%, GPQA Diamond 84.5%, Humanity's Last Exam — with tools 44.9%, Humanity's Last Exam — without tools 26.9%, LiveCodeBench 83.1%, SWE-bench Verified 71.3%, Terminal-Bench 35.7%, FrontierMath 21.4%.
Capabilities
- Extended reasoning for mathematical, scientific and analytical text tasks.
- Code generation and repository-level software engineering with appropriate tools.
- Interleaved reasoning and tool calling across long agent workflows.
- Processing long documents and code within a documented 256K-token context window.
- Open-weight deployment with a published INT4 quantized format.
What improved
- Introduces a dedicated thinking model alongside the earlier Kimi K2 instruction-following models.
- Expands context from the original K2 release's 128K tokens to 256K tokens; this is not an increase over every intervening K2 variant.
- Emphasizes sustained reasoning interleaved with long sequences of tool calls.
- Uses quantization-aware training for the published INT4 deployment format.
Known flaws & limitations
- Text-only inputs do not provide native image, audio or video understanding.
- The trillion-parameter mixture-of-experts architecture requires substantial memory and serving infrastructure despite sparse activation and quantization.
- Extended reasoning and repeated tool calls can increase latency and total token expenditure.
- Tool-assisted benchmark results depend on external tools, evaluation scaffolding and execution budgets.
- Strong benchmark performance does not guarantee factual accuracy or successful completion of unfamiliar tasks.
Risks, safety & security issues
- Agent deployments with excessive tool permissions can expose files, credentials and external services; use least-privilege access.
- Untrusted documents and tool outputs can introduce prompt-injection instructions; isolate them from privileged actions.
- Generated code and shell commands can be unsafe or incorrect; require sandboxing and review before execution.
- Sensitive information included in prompts or tool results can propagate into logs or downstream services; minimize and redact sensitive data.
Good fit for
- Complex text-based reasoning where additional inference time is acceptable.
- Coding and repository maintenance with tests, sandboxed execution and human review.
- Long-document analysis and synthesis.
- Tool-assisted research and multi-step workflows with tightly scoped permissions.
Poor fit for
- Native image, audio or video understanding.
- Latency-critical applications that require consistently short responses.
- Small-device or low-memory self-hosting.
- Unsupervised consequential decisions or unrestricted access to production systems.
References
Profile last reviewed 10/20/2026 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.