Moonshot AI releases Kimi K2 Thinking on Thursday, an open-weights reasoning model the Beijing startup describes as its most capable to date. It arrives four months after July’s K2. The model card lists a 1-trillion-parameter mixture-of-experts architecture with 32 billion active parameters, a 256K context window, native INT4 quantization, and training that interleaves chain-of-thought reasoning with tool calls.

Moonshot claims the model maintains coherent goal-directed behavior across 200 to 300 consecutive tool invocations, where prior models degrade after 30 to 50. On BrowseComp with tools, it reports 60.2% to GPT-5’s 54.9%; on Humanity’s Last Exam with tools, it reports 44.9%; the footnote says GPT-5 scores 41.7 on Moonshot’s internal HLE text-only subset. A caveat sits in the footnotes: with Hugging Face access unblocked, K2 Thinking scores 51.3 on HLE, so the published number reflects blocked access.

A source familiar with the matter tells CNBC the training run cost $4.6 million. The outlet could not independently verify the figure.

The release lands Thursday, as Nvidia CEO Jensen Huang this week again urges the U.S. to press ahead in the race against Chinese-developed AI, and some major U.S. companies, including Airbnb, have begun to publicly tout Chinese models as viable — and often cheaper — alternatives to OpenAI’s.