Moonshot releases open-source Kimi K2 Thinking
Moonshot AI's Kimi K2 Thinking is an open-source reasoning agent that can run 200–300 tool calls in a row, and the company reports record scores on agentic benchmarks such as HLE and BrowseComp.
Moonshot AI has introduced Kimi K2 Thinking, which it calls its best open-source thinking model. The model is built as an agent that reasons step by step while it uses tools. Moonshot says it can make 200–300 sequential tool calls without human input. The company describes this as part of its test-time scaling work, which scales both thinking tokens and tool-calling steps.
Moonshot reports these results:
- Humanity's Last Exam (with search, Python and browsing tools): 44.9%
- BrowseComp: 60.2%. The human baseline is 29.2%.
- SWE-Bench Verified: 71.3%
- SWE-Multilingual: 61.1%
- Terminal-Bench: 47.1%
The company calls several of these state of the art and says the model "matches or surpasses" the latest open-source and frontier models. These are Moonshot's own numbers and have not been independently reproduced. A footnote in the post says the consumer chat product uses fewer tools and a lower tool-call limit than the full agent.
On efficiency, Moonshot says it used quantization-aware training after the main training run. It applied INT4 weight-only quantization to the model's mixture-of-experts layers. The company says this gives native INT4 inference with roughly twice the generation speed, and that all reported benchmarks were run at INT4 precision. That addresses a known problem: thinking models produce long outputs, and quantization often causes large accuracy drops in them.
K2 Thinking is live now in chat mode on kimi.com and through the Kimi K2 Thinking API. The full agentic mode is "available soon." The post also claims improvements in creative and practical writing, but it gives no figures for those.