Who is Moonshot AI and why you don't know them

Moonshot AI was founded in March 2023 in Beijing and belongs to the so-called "Big Six of Chinese AI tigers" — an elite group of startups dominating the country's AI development. It was founded by three classmates from the prestigious Tsinghua University: Yang Zhilin (CEO), Zhou Xinyu, and Wu Yuxin. The company name — "the dark side of the Moon" — references the famous Pink Floyd album, released exactly 50 years before the company was founded. Unlike Western AI giants such as OpenAI or Anthropic, which employ thousands of people, Moonshot AI operates with just 300 employees. And what's even more surprising — the company rejected traditional corporate hierarchy. According to available information, it has no conventional management layers and relies on a flat organizational structure where researchers work directly, without intermediaries. This approach is absolutely exceptional in China's corporate environment — and it's clearly working so far. In less than three years of existence, the startup has raised investments exceeding $1.9 billion (over 40 billion CZK). The main investor is Alibaba Group, which holds a 36% stake, with additional funding from Tencent, Gaorong Capital, and IDG Capital. In March 2026, Moonshot began considering an IPO in Hong Kong, which — according to Forbes estimates — could make Yang Zhilin another in a series of Chinese billionaires who made their fortunes from artificial intelligence.

Kimi: From unknown chatbot to market leader

The first version of the Kimi chatbot saw the light of day back in October 2023. Its key advantage was the ability to process up to 200,000 Chinese characters at once — roughly 150 standard pages of text. At that time, Kimi gained attention primarily in the domestic market and quickly became one of the strongest competitors to Baidu Ernie Bot. The real breakthrough, however, came in 2025. In January, Moonshot released the Kimi K1.5 model, which the company claimed matched OpenAI o1's performance in math, programming, and multimodal reasoning. This was followed by Kimi K2 Thinking (November 2025) — an open model with a trillion parameters on a Mixture-of-Experts (MoE) architecture, which surpassed GPT-5 and Claude Sonnet 4.5 in key tests: on the Humanity's Last Exam benchmark it scored 44.9%, in BrowseComp 60.2%, and in SWE-Bench Verified 71.3%. Training the model cost just $4.6 million — a fraction of the cost of comparable American models. In January 2026, multimodality arrived — the Kimi K2.5 model received a 400-million-parameter visual encoder MoonViT, enabling work with images and video. And now, on July 16, 2026, the pinnacle arrives: the Kimi K3 model with 2.8 trillion parameters.

Kimi K3: What it can do and how it stacks up against Western competition

Kimi K3 is the first open model to break the 2-trillion-parameter barrier. It uses an advanced MoE architecture, where only a portion of parameters is activated at runtime — meaning high performance at relatively low operational costs. The result is a price comparable to the cheaper Claude Sonnet model, yet performance that, according to independent benchmarks, surpasses Claude Opus 4.8 max and GPT-5.5 high. It still trails the cutting edge in the form of Claude Fable 5 and GPT-5.6 Sol, but the gap is narrowing dramatically. What can K3 do? It is designed for long-duration coding tasks, knowledge work, and complex reasoning. It can optimize GPU kernels, produce physics research results, edit video, and — what's particularly remarkable — it edited its own promo video from 56 source clips, including shot selection, syncing edits with music, and multiple rounds of revisions. On an independent private evaluation of long-term knowledge work, K3 achieved an overall Elo score of 1,547 points, which is 732 points more than the previous K2.6 version. It left all models in the dust except Claude Fable 5.

An open approach that changes the rules of the game

Moonshot AI is among the companies that have bet on an open-source strategy. The Kimi K2, K2 Thinking, and K3 models are released under a modified MIT license that allows free use — with the exception of products with more than 100 million monthly users or revenues over $20 million per month. In other words: any developer can download the model, modify it, and run it on their own hardware. This is a crucial advantage over OpenAI's closed models, which are only available via API. Upon the release of Kimi K2, the model became the most downloaded model on Hugging Face within 24 hours — an unprecedented success for a Chinese model. The open approach simultaneously responds to U.S. efforts to curb China's technological development — when the West says "we won't supply you chips," China counters with: "here's our best model for free." The main Kimi chatbot is available free of charge at kimi.com, with paid plans starting at just a few dollars. The model also supports the Czech language — although it primarily targets the Chinese market, thanks to its enormous context window (up to 256,000 tokens) it handles Czech input without issues. For Czech users and businesses, this represents an interesting alternative to Western chatbots — especially for working with longer documents, where Kimi excels.

The era of Chinese AI: It's no longer just about DeepSeek

Last year's success of the DeepSeek R1 model showed that Chinese AI labs can compete with the American elite at a fraction of the cost. This year, Moonshot AI adds another dimension: a flat organizational structure, open-source philosophy, and cost efficiency that is unsettling investors in Silicon Valley. Controversies, however, have not bypassed Moonshot. In February 2026, Anthropic accused Moonshot (along with other Chinese companies) of violating terms of service — allegedly using thousands of fraudulent accounts to generate millions of conversations with Claude for training their own models. Moonshot did not comment on the allegation. Wherever the truth lies, it shows how intensely the fight for AI dominance is intensifying — and that the line between "inspiration" and "theft" is getting ever thinner. For the average user, however, something else is key: competition drives prices down and quality up. Just two years ago, a top-tier AI model was the privilege of a handful of companies. Today, you can run one for free, in Czech, from a computer in Prague. And that's partly thanks to a company you may never have heard of.

Practical impact: What Kimi K3 means for the Czech Republic and Europe

Kimi K3 is an open model — meaning that Czech companies and developers can use it without paying license fees. For the startup scene, which often lacks the budget for expensive API services from OpenAI or Anthropic, this is a significant opportunity. The model runs on Moonshot's own infrastructure (the so-called Mooncake platform, which processes 100 billion tokens daily and whose design won a prestigious award at the USENIX FAST conference), but thanks to open weights, it can also be run locally — if you have sufficiently powerful hardware. For European companies dealing with GDPR and data sovereignty, this is a strong argument. In the context of the EU AI Act, it is not yet clear exactly how Chinese models will be regulated — Moonshot has no European branch and does not officially target the European market. Nevertheless, its model is freely available and nothing prevents its use for non-commercial and commercial purposes within the licensing terms.