Moonshot AI releases Kimi K2 Thinking

Background

Kimi K2 is an open-weight language model with 1 trillion total parameters and a mixture-of-experts layout. In November 2025 Moonshot AI released Kimi K2 Thinking, a reasoning version of it ^[1][2]. Like other reasoning models, it produces intermediate reasoning before its final answer, which is why the length of its tool-use chains and its speed both matter.

A reasoning model is one that works through a problem before answering. Reasoning models have become important because many useful tasks, from mathematics to multi-step coding, depend on a sequence of intermediate steps rather than a single guess.

What happened

In November 2025 Moonshot AI released Kimi K2 Thinking as a reasoning model ^[1][2]. The model keeps the mixture-of-experts layout of Kimi K2: 1 trillion total parameters, 32 billion active per token, 384 experts with 8 selected and 1 shared, and a 256K-token context window ^[3].

The training cost was reported at about US$4.6 million ^[1].

Key details

The release used quantization-aware training in post-training to create native INT4 weight-only quantization for the mixture-of-experts components. Moonshot says this roughly doubles generation speed ^[3].

A second key detail is tool use. Moonshot says the model keeps stable tool use across 200 to 300 sequential calls, while earlier models degrade after 30 to 50 steps ^[3]. A sequential call here means one tool invocation after another, where each result feeds the next step, so stability over a longer run decides whether a multi-step task finishes.

On benchmarks, Moonshot reported 44.9 on Humanity's Last Exam with tools, 51.0 in Heavy mode, 60.2 on BrowseComp and 71.3 on SWE-bench Verified ^[3]. These are the company's reported results.

Impact

The release combined an open-weight reasoning model with tool use over long sequences of calls, which matters for agents that work through many steps. The quantization of the mixture-of-experts parts is aimed at generation speed, which Moonshot says roughly doubles ^[3].