Kimi K2 Thinking

What it is

Kimi K2 Thinking is a reasoning model released by Moonshot AI in November 2025^[1][2]. A reasoning model is one trained to work through intermediate steps — checking its own work, trying an approach and abandoning it, calling tools — before it produces a final answer, instead of writing a reply in a single pass. Kimi K2 Thinking is a direct descendant of Kimi K2, and Moonshot kept the earlier model's architecture: a mixture-of-experts (MoE) layout with 1 trillion total parameters, 32B active at a time, drawn from 384 experts of which 8 are selected for each token plus 1 shared expert. The context window is 256K tokens^[3].

A mixture of experts is a sparse design: the network contains many separate sub-networks, the "experts", but only a few of them are used for any given piece of text, so the cost of a single forward pass stays far below what a dense model of the same total size would need. That is how a model can carry a trillion parameters while activating a small fraction of them per token.

Native low precision

One of the more distinctive engineering choices is how the model is stored and served. During post-training, Moonshot applied quantization-aware training so that the mixture-of-experts components could be quantized natively to INT4 weight-only precision^[3]. In plain terms, the model was trained while already subject to the rounding error it would face at inference time, rather than being squeezed down to low precision afterwards. Moonshot says this roughly doubles generation speed^[3]. The model card also states that all benchmark results were reported under INT4 precision^[3], so the published numbers describe the quantized model that users actually run, not a higher-precision reference version.

Long tool chains

Moonshot highlights the model's behaviour when it is used as the driver of an agent that calls external tools. According to the company, the model maintains stable tool use across 200 to 300 sequential calls, whereas earlier models degrade after 30 to 50 steps^[3]. Long tool chains matter because tasks such as searching the web across many pages, editing a code repository, or running a terminal session require the model to keep track of what it has already done across dozens or hundreds of intermediate actions.

Reported results

Several benchmark figures were published with the release. On Humanity's Last Exam, a broad set of expert-level questions, the model scored 44.9 with tools, and 51.0 in a configuration labelled Heavy mode^[3]. On BrowseComp, which tests finding information by browsing, it scored 60.2; on SWE-bench Verified, which measures fixing real software issues drawn from code repositories, it scored 71.3^[3]. Without tools it scored 94.5 on AIME25 and 83.1 on LiveCodeBench v6^[3]. The model card also lists GPQA at 84.5 without tools and Terminal-Bench at 47.1 with simulated tools^[3].

Cost and licence

The training cost was reported at about US$4.6 million^[1]. The model was released under a modified MIT licence, which Wikipedia notes requires attribution for products with more than 100 million monthly users or US$20 million in monthly revenue^[3][1].

Where it sits

Kimi K2 Thinking was succeeded by Kimi K2.5 in January 2026^[1]. Its place in the line is that of the reasoning-focused member of the K2 family: the same base architecture as K2, retrained for step-by-step problem solving, long tool use, and efficient low-precision serving.