Kimi K2, K2 Thinking and K3 in Detail: Moonshot AI’s Model Series Explained

Anyone who looks more closely at Moonshot AI’s model series quickly comes across a whole series of names: Kimi K2, Kimi K2 Thinking, K2.5, K2.6, K2.7 Code and finally K3. Unlike many Western providers, who rarely disclose their model versions, Moonshot AI publishes technical reports for every major version, along with the model weights themselves for download. For anyone interested from the real estate and finance industry, a closer look is worthwhile, because behind the version numbers lie concrete differences in context window, computational cost and agentic capabilities – and therefore also in the question of what a model is actually suited for in everyday work.

From research model to heavyweight: the Kimi K2 series

Kimi K2: the technical foundation

The original Kimi K2 version was introduced in July 2025 and described in detail in a technical report (“Kimi K2: Open Agentic Intelligence”). Technically, it is a so-called Mixture-of-Experts (MoE) model: instead of a single, continuously active neural network, the model consists of many specialized sub-networks (“experts”), of which only a small portion is actually activated per request. For Kimi K2, according to Moonshot AI, there are a total of around one trillion parameters, of which around 32 billion are actively used for any single request. The model is spread across 384 experts, of which 8 are specifically selected per text segment, supplemented by an always-active “shared” expert.

This structure explains why Kimi K2, despite its enormous overall size, can be operated comparatively economically: the entire trillion parameters do not need to be computed for every request, only a small, selected part. According to Moonshot AI, the model was trained using a specially developed optimization method (MuonClip), intended to improve training stability for such large MoE models. The original base version offered a context window of 128,000 tokens; the Instruct version (for dialogue and instructions) released at the same time was expanded to 256,000 tokens.

Kimi K2 Thinking: thinking and acting in one model

In autumn 2025, according to consistent reports from November 2025, Moonshot AI introduced “Kimi K2 Thinking,” a variant that additionally incorporates extensive, multi-step thinking before the actual answer and, while still in the thinking process, can already call on tools (such as a web search or a calculator). According to available benchmark figures, the model achieved a score of around 71 percent on the well-known programming test “SWE-Bench Verified” and, in test scenarios, was able to independently carry out several hundred consecutive tool calls to solve complex, lengthy tasks. This deliberately positioned Moonshot AI in the field of so-called agent models, which don’t just answer but independently plan and execute multi-step tasks.

Further development in 2026: K2.5, K2.6 and K3

Over the course of 2026, Moonshot AI released several further tiers. According to consistent, though not always uniform, sources, K2.5 appeared, followed in April 2026 by K2.6 with a context window expanded to 256,000 tokens and improved programming capabilities, followed by a variant specialized in programming tasks called K2.7 Code. Finally, in July 2026, Kimi K3 followed, which, according to the manufacturer, comes with a context window enlarged to around 1 million tokens and native image processing (vision) – according to Moonshot AI, four times the size of K2.6’s context window. Various sources give differing figures for K3’s exact parameter size (some in the range of roughly two to three trillion total parameters); a single, officially definitive figure cannot currently be confirmed with complete certainty, which is why we deliberately speak of an order of magnitude rather than an exact number here.

What makes the Kimi models technically and strategically distinctive?

Open weights with a commercial condition

Moonshot AI releases the model weights for K2 and its successors under a self-formulated, modified MIT license. At its core, this is a very permissive open-source license that also allows commercial use. One special feature: anyone operating a product or service based on the models with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must visibly display the notice “Kimi K2” in their own interface. For the vast majority of smaller and medium-sized use cases – including in the real estate industry – this clause practically never applies, but it shows that “open” at Kimi is not meant entirely unconditionally.

Agentic capabilities and a large context window

The strategic focus of the Kimi models is clearly on so-called agentic capabilities: the ability to independently handle multi-step tasks with tool use, web access and code execution, rather than just answering individual questions. Combined with a very large context window – around one million tokens for K3 – this makes it possible to process very extensive document collections in a single pass, such as several hundred pages of contract documents or a larger collection of property documents.

Rule of thumb: the large context window makes it possible to read in a great many documents at once – but this does not replace careful review of the output, since even a large context window does not protect against errors or omissions in the result.

Practical examples of use

1. Automated summarizing of large file collections

A case worker at a property management company could have several hundred utility bill statements or minutes from owners’ meetings read in via the API in a single pass, in order to automatically flag anomalies – such as unusually high cost items. The large context window of K2.6 or K3 makes this technically possible without first having to split the documents into many small chunks.

2. Drafting and laying out a listing exposé

In agent mode, a real estate agent can have a first, multi-page draft version of a property exposé created, including a simple web presentation, which is then reviewed editorially by hand and cross-checked against the actual, legally binding property data.

3. Programming support for internal tools

Since the Kimi models perform particularly strongly on programming benchmarks, some IT departments at real estate companies use Kimi to help build small internal automation scripts, for example to reconcile rent lists between two software systems.

4. Research and market monitoring

Via the web-browsing feature of agent mode, a model can be used to continuously monitor publicly accessible market data, for example to summarize interest rate trends or publicly accessible real estate market reports – it’s important here to work exclusively with publicly available, non-personal sources.

5. Multilingual communication

For international real estate portfolios, a financial case worker can use Kimi to translate standard letters or summaries between several languages – though the restriction on personal content described above still applies here as well.

Rule of thumb: the larger the volume of data a model processes in one pass, the more important the question becomes of exactly where that data resides during processing – performance and data protection are two separate questions, and both need to be answered.

FAQ about the Kimi models

What is the difference between Kimi K2 and Kimi K2 Thinking?

Kimi K2 is the original base model; Kimi K2 Thinking is a variant introduced in November 2025 that incorporates more extensive, multi-step thinking with tool use before answering, and thereby performs better on complex tasks.

How large is the context window of the current models?

The models in the K2 series generally offer 256,000 tokens of context; Kimi K3, introduced in July 2026, is stated to have around one million tokens.

Are the Kimi models really open source?

The model weights are openly accessible and released under a modified, very permissive MIT license. One special feature is the attribution requirement for very large commercial applications (over 100 million users or 20 million US dollars in monthly revenue).

Are the models suited to programming tasks?

Yes, the Kimi K2 series, and especially the specialized K2.7 Code variant, perform comparatively strongly on public programming benchmarks and are frequently used for coding agents.

Can I run the models myself instead of using the Kimi website?

Yes, since the weights are released openly, they can be run via various cloud providers or on your own hardware. However, this requires technical expertise and sufficient computing capacity, since these are very large models.

Basic information about the company Moonshot AI, pricing and data protection can be found on our provider page. A comparison of all ten AI providers in the cluster is available on the overview page.