Moonshot AI Kimi — China's Math and Coding AI Startup

The Fastest Chinese Company to Reach $10 Billion Valuation in History

In less than two weeks since launching its K2.5 model, Moonshot AI’s revenue exceeded its total revenue for all of 2025. This sentence alone summarizes what happened to this Chinese startup in early 2026: a rocket that suddenly rose. From a $4.3 billion valuation at the end of 2025, to $10-12 billion in February 2026, to negotiations to reach $18 billion in March 2026.

Founding Story: A Tsinghua Professor Builds an OpenAI Competitor

Moonshot AI was founded by Yang Zhilin in 2023 in Beijing. Yang was a former assistant professor at Tsinghua University, taught at American universities, and worked on AI projects at Meta and Google. The decision to spin off and launch his own company came with the wave of great enthusiasm unleashed by ChatGPT in 2022.

Yang chose a specific and distinctive specialization: models with ultra-long context. While most models processed thousands of tokens, he designed Kimi to work with hundreds of thousands of Chinese characters in a single conversation — a feature that Chinese users discovered they desperately needed for analyzing long documents, legal reports, and business documents.

The name “Kimi” is taken from Yang’s shortened English name.

Main Products and Models

Kimi (October 2023)

The company’s first AI assistant. At its launch, it could process 200,000 Chinese characters in a single conversation — a unique feature at the time. It quickly became the third most popular assistant in China in terms of active users.

Kimi (March 2024 — Extended Context)

Yang claimed that Kimi could now process 2 million Chinese characters in a single conversation — a massive leap that made Kimi comparable to Claude in long-context capabilities. The announcement was so huge that it caused service outages for two days due to the influx of users.

Kimi K1.5 (January 2025)

A reasoning model. The company claimed it matched OpenAI o1’s performance in mathematics, programming, and multimodal reasoning.

Kimi K2 (July 2025)

A massive open-source model with one trillion total parameters (32 billion active) with MoE architecture — one of the largest open-weight models in the world at launch. Trained on 15.5 trillion tokens and licensed under a modified MIT license. On its launch day, it ranked first in downloads on Hugging Face. Excels in programming tasks and passes LiveCodeBench.

Kimi K2-Instruct-0905 (September 2025)

An update that doubles the context window to 256,000 tokens and improves Agentic programming capabilities.

Kimi K2 Thinking (November 2025)

A version optimized for deep thinking and multi-step planning. Training cost: only ~$4.6 million. Capable of executing 200-300 automatic tool calls in a single sequence.

Kimi K2.5 (January 2026)

The real bombshell. Added comprehensive visual capabilities through a visual encoder with 400 million parameters called MoonViT, enabling the model to process images and video and perform foreign execution (such as reproducing a user’s journey on a website by watching a video describing that journey). In less than 20 days from launch, its revenue exceeded the company’s total revenue throughout 2025.

Mooncake Platform

The infrastructure serving Kimi processes 100 billion tokens daily. Received the Erik Riedel Best Paper Award at the USENIX FAST conference for its architectural design.

Achievements and Numbers

The funding trajectory resembles a rocket in shape:

Date	Round	Amount	Valuation
2023	Seed	$300 million	$2.5 billion
February 2024	B	Raised to $2.5B	—
October-December 2024	C	$500 million	$4.3 billion
February 2026	New Round	$700 million (led by Alibaba and Tencent)	$10-12 billion
March 2026	Expansion	Negotiations	$18 billion

Total Announced Funding: Over $1.2 billion.

Record Speed: From seed round to $10 billion valuation in less than two years — the fastest Chinese company to achieve this number in history, surpassing ByteDance (four years) and Pinduoduo (three years).

International Growth: International revenue exceeded domestic revenue in early 2026, with Kimi K2.5 topping global OpenRouter rankings.

Technical Performance: K2.5 was classified by Artificial Analysis as “among the strongest open models globally,” with a slight gap behind the latest Western models from OpenAI and Anthropic on most benchmarks.

Competition and Challenges

Market Position in China

At the peak of its popularity, Kimi was the third most used assistant in China after Baidu Ernie and Alibaba Qwen. But with the crowding of dozens of Chinese models, it slipped to seventh place by June 2025 — before K2 and K2.5 brought it back to the forefront.

External Concerns

Allegations from Anthropic (and other companies) that Moonshot and its Chinese peers used “distillation” techniques to extract outputs from Western models to improve their own — an approach controversial ethically and legally.

US chip export restrictions hinder Moonshot’s access to Nvidia H100/H200 processors, forcing it to improve model efficiency to run on less capable hardware.

Future Vision 2026–2027

Hong Kong IPO: In March 2026, reports revealed that Moonshot is considering an IPO on the Hong Kong Stock Exchange. Yang Zhilin explicitly stated: “no plans for IPO in the near term” — but financial advisors began preparations according to sources.

Global Expansion: International revenue exceeds domestic for the first time — an indicator that Moonshot is transitioning from a Chinese company to a global player.

Kimi Claw: The advanced agent platform that exploded demand in early 2026 — investment in improving it represents the top priority.

Analytical Conclusion

From few indicators in 2023 — long context and a polished Chinese assistant — Moonshot built one of the most exciting growth stories in global AI in 2026. The valuation from $2.5 billion to $18 billion in less than two years reflects investor confidence, but also raises expectations to levels difficult to sustain continuously.

The business model is clear and logical: subscriptions for individual users, and API fees for developers and enterprises. The challenge is sustainability in a market where DeepSeek, Ali Qwen, and others offer free or low-cost models. If K2.5 succeeds in maintaining its global growth, it may prove that Chinese open-source AI is capable of competing globally without needing government contracts or massive enterprise subscriptions.