January marked the end of a frenetic quarter of major large model releases. For Moonshot AI’s Kimi, which had just rolled out its K2.5 model, the period became a clear inflection point.
In recent discussions with investors, Moonshot said its overseas revenue has surpassed its domestic revenue. Following the release of K2.5, the company’s global paying user base grew fourfold. That shift, according to the company, took place within days of the model’s debut.
Interest from outside China has been building since the release of K2. After K2.5 launched, that momentum accelerated. On OpenRouter, K2.5 ranked third as of publication, trailing only Claude Sonnet 4.5 and Gemini 3 Flash.
Commercialization also moved quickly. Moonshot began monetizing Kimi K2 in October, a relatively short timeline by industry standards.
In an internal letter circulated late last year, founder Yang Zhilin wrote that since November 2025, Kimi’s overseas API revenue had quadrupled. Monthly growth rates for both overseas and domestic paying users exceeded 170%, according to the letter.
On January 27, Moonshot released and open-sourced Kimi K2.5, which it described as its most capable model to date. Built on a native multimodal architecture, K2.5 supports visual understanding, code generation, agent clusters, and both reasoning and non-reasoning modes.
Across benchmarks including Humanity’s Last Exam, BrowseComp, and SWE-Bench Verified, Kimi said K2.5 delivered robust performance among open-source models. On some metrics, the company claims the model outperformed closed-source systems such as GPT-5.2 and Claude Opus 4.5.
If DeepSeek used R1 to show that Chinese large language models could compete on reasoning performance, Moonshot is pursuing a different path, emphasizing the collaborative capabilities of artificial intelligence systems.
Over the past year, Kimi’s model iterations from K1.5 to K2.5 have followed a consistent trajectory, pushing the system toward functioning as an autonomous agent rather than a conventional chatbot.
During the K1.5 phase, the focus was on long-context understanding and generation. K2 marked a scaling stage, centered on agent task execution and more complex operations. With K2.5, the company is testing how coordinated agent clusters can operate through what it describes as “teamwork.”
In real-world deployments, K2.5 can reportedly orchestrate up to 100 agents and run as many as 1,500 steps in parallel. Kimi said this has expanded the practical usefulness of agents. In large-scale information-gathering scenarios, agent clusters have improved efficiency by three to ten times, according to internal measurements.
Typical applications include collecting and summarizing recent research on a given topic, organizing outputs into Excel files, extracting interface logic from screen recordings and generating frontend code, and automating office workflows such as editing Word documents, building Excel data models, creating PowerPoint presentations, and translating or revising PDFs.
Yang addressed the rationale behind this approach during a Reddit ask-me-anything session on January 29.
He said the growth rate of high-quality data can no longer keep pace with the growth of compute. As a result, the traditional scaling approach of predicting the next token using internet data delivers diminishing returns. Alternative approaches, Yang said, include agent swarms, which scale performance by increasing the number of agents executing subtasks in parallel. He described this as a form of test-time scaling that could also support training-time scaling.
Following the release of K2.5, Kimi’s trajectory has begun to resemble a hybrid of Anthropic’s research-driven model development and Manus’ agent-oriented product philosophy.
On the model side, Kimi has positioned itself against Anthropic, prioritizing foundational model capability while building technical influence through open source. Starting with K2, Moonshot fully open-sourced its Kimi model weights and toolchain, allowing developers to deploy the models locally or in the cloud.
Moonshot has maintained a relatively lean organization. Its team numbers about 300 people, roughly one-tenth the size of many large technology companies. Internally, the company often highlights a stark comparison: delivering globally competitive models while using about 1% of the compute resources available to larger peers.
This emphasis on efficiency places algorithms and engineering decisions at the center of its strategy. Limited resources have also narrowed the company’s focus to a small number of forward-looking bets. Past examples include becoming the first team to run the Muon optimizer at scale in LLM training and developing an in-house linear attention mechanism, both cited internally as milestones.
On the product side, Kimi has settled into a clearer structure. One arm serves developers through its Kimi API platform, targeting a global developer audience. The other focuses on consumer products, where Kimi is positioning itself primarily as a productivity tool.
The company has been consolidating its branding while making its products more general-purpose. An earlier consumer-facing agent product, OK Computer, has been renamed Kimi Agent in recent updates.
Across use cases, Moonshot has continued refining polish and usability, adapting interaction styles to different contexts. One area of emphasis is editability in complex workflows. After generating PowerPoint slides or animations from Excel data, for example, users can break elements apart and modify them directly within Kimi, a capability that depends heavily on model performance.
In early December, Moonshot president Zhang Yutong said the company needed to find its own narrative and identify what it truly excels at.
“When competing with much larger companies that have more resources, we deliberately control our business boundaries,” Zhang said. “We focus on the large-model layer, the logic layer, the agent layer, and productivity-oriented, complex task workflows such as in-depth research, presentations, data analysis, and website development.”
Across the sector, model developers are converging on similar agent-driven scenarios, particularly in coding and office productivity. These use cases offer relatively clear paths to commercialization but also depend on continued gains in model capability.
For Moonshot, the stated ambition is to build a “platform agent.” Achieving that will require sustaining top-tier model performance while also developing consumer products with a distinct identity and consistent design sensibility.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Deng Yongyi for 36Kr.