Kimi K2 Thinking

Kimi K2 Thinking, released November 6, 2025, is Moonshot's agentic reasoning variant, extending the Kimi K2 foundation with deliberate multi-step reasoning and extended tool coordination. It is built on the 1 trillion total parameter MoE architecture with 32 billion active parameters and is tuned for complex tasks that benefit from explicit step-by-step planning. The model can execute 200 to 300 sequential tool calls without human intervention, supporting autonomous work across intricate, multi-stage workflows. It reports strong results on reasoning-intensive benchmarks, including 44.9% on Humanity's Last Exam with tools, 60.2% on BrowseComp, and 71.3% on SWE-Bench Verified, competing with frontier proprietary models. Its core strength is planning before answering, breaking tasks into sub-goals, explaining reasoning steps, and coordinating tools when useful. It supports a 256K token context window and is released as open source under a Modified MIT License.