Introduction
While OpenAI’s ChatGPT remains unavailable in mainland China, a Beijing-based startup has spent the past three years building what’s become one of the country’s most-used AI chatbots in its place. Kimi, developed by Moonshot AI, first launched in October 2023 and quickly became one of the largest AI services in China by user base. This month, the company raised the stakes significantly: its newest model, Kimi K3, is now claiming performance that rivals — and in some specific benchmarks, beats — the leading systems from both OpenAI and Anthropic.
Here’s who’s behind Kimi, how the company got here, and what its latest release actually claims to do.
Table of Contents
- Who Is Moonshot AI?
- How Kimi Became China’s ChatGPT Alternative
- What Kimi K3 Actually Claims
- How K3 Performs on Independent Benchmarks
- Why Open-Weight Matters Here
- The Market Reaction
- Where Kimi Fits in the Broader US-China AI Race
- Conclusion
Who Is Moonshot AI?
Moonshot AI is a Beijing-based AI company founded in 2023, and it’s counted among what’s sometimes referred to as China’s “AI Six Dragons” — a group of the country’s leading AI model developers. The company has drawn backing from two of China’s largest tech giants, Alibaba and Tencent, and at one point held the distinction of being the highest-valued large language model company in China, with a funding round last year valuing the company at roughly $3 billion. More recently, the company raised approximately $2 billion in additional funding as it pushed toward its K3 release.
How Kimi Became China’s ChatGPT Alternative
Kimi’s rise has been built on two consistent strategies: aggressive product features and low pricing. Early versions of the chatbot stood out for supporting lossless context input of up to 2 million characters — letting users feed in enormous amounts of text at once — combined with large-scale promotional activity that helped Kimi become one of the largest AI services in China by user base.
The company has also leaned into open-source releases. Its Kimi K2 model, launched last year, was positioned as a direct, lower-cost competitor to OpenAI’s ChatGPT and Anthropic’s Claude, particularly in code generation, and was made available for developers to use freely under a license requiring attribution only from extremely large commercial deployments. A subsequent update, Kimi K2 Thinking, focused specifically on agentic capability — the ability to understand what a user wants and act on it without requiring explicit, step-by-step instructions — and was released as Nvidia CEO Jensen Huang was publicly urging the US to keep pressing its lead over Chinese AI development.
More recently, Moonshot has also been testing community and social features built directly into the Kimi app, including a feed where users can share their AI interactions with each other — a strategy read by industry observers as an attempt to increase user engagement and compete for attention against rival Chinese chatbot DeepSeek.
What Kimi K3 Actually Claims
Moonshot introduced Kimi K3 as what it describes as the world’s largest open AI model, built with roughly 2.8 trillion parameters — a figure the company disclosed even though neither OpenAI nor Anthropic has published equivalent parameter counts for their own current models, making direct size comparisons difficult to verify independently.
According to Moonshot, K3 still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall performance, but consistently outperformed every other model the company tested it against. On specific benchmarks involving coding and general agentic tasks, Moonshot said K3 performed competitively with Claude Fable 5 and substantially outperformed Claude Opus 4.8, along with OpenAI’s Sol and GPT-5.5 models — systems that sit just behind each company’s current leading-edge releases.
How K3 Performs on Independent Benchmarks {#how-k3-performs-on-independent-benchmarks}
Independent testing has offered some corroboration of Moonshot’s claims. On the AI benchmark platform Arena.ai, Kimi K3 jumped from 17th to 1st place in the “Frontend Code Arena” category, which measures a model’s ability to perform multi-step web development tasks — surpassing both Claude Fable 5 and GPT-5.6 Sol by a wide margin on that specific benchmark. In the platform’s “Text Arena” category, which measures text-to-text tasks like creative writing, K3 reached 9th place, a substantial jump from the 38th-place ranking held by Moonshot’s previous model, Kimi K2.6.
Bank of America analysts also weighed in positively on the release, noting that despite Chinese AI firms generally having more limited access to the most advanced chips used for training, Moonshot has demonstrated it can still make major performance advancements through improvements in how it trains and designs its models rather than through raw hardware advantage alone.
Why Open-Weight Matters Here {#why-open-weight-matters-here}
A key part of K3’s positioning is that it’s an open-weight model, meaning developers can download, run, and modify the underlying system directly — a structural difference from proprietary systems like OpenAI’s ChatGPT and Anthropic’s Claude, where the underlying models remain private and are only accessible through the companies’ own products and APIs. Industry commentary has drawn comparisons between K3’s release and DeepSeek’s disruptive debut earlier in the AI race, given that K3 is closing in on frontier-level performance on specific benchmarks while reportedly being offered at a fraction of the cost of leading proprietary alternatives.
The Market Reaction
K3’s release had an immediate, visible impact on Moonshot’s domestic competitors. Shares of two rival Chinese AI companies fell sharply on the day of the announcement: Zhipu dropped 28.4% and MiniMax fell 15.6% in Hong Kong-based trading, while shares of Z.ai — which had released its own new model just the month before — dropped by an identical 28.4%. That kind of same-day reaction across multiple competitors suggests investors read K3’s release as a meaningful competitive shift within China’s own crowded AI landscape, not just a story about catching up to US labs.
Where Kimi Fits in the Broader US-China AI Race
K3’s release lands squarely inside an intensifying broader competition between American and Chinese AI developers. Some industry analysts have pushed back on characterizing the reaction to K3 as overblown, instead attributing much of the intensity to an active political debate in Washington DC over how the US should respond to China’s continued AI progress.
It’s also worth noting that Kimi’s rise as a consumer product happened specifically because ChatGPT has never been officially available inside mainland China — leaving an opening that Moonshot, alongside other domestic players, has spent the past three years filling. K3’s benchmark performance suggests that opening is now translating into genuine technical competitiveness on the world stage, not just domestic market share built on ChatGPT’s absence.
Conclusion
Kimi’s path from a 2023 startup product to a model now drawing direct benchmark comparisons with Anthropic and OpenAI’s leading systems reflects a broader pattern playing out across China’s AI industry: rapid iteration, aggressive open-source releases, and a willingness to compete hard on price even with more limited access to advanced training hardware. Moonshot’s own framing is notably measured — K3 still trails the very top US models on overall performance, by the company’s own account. But the specific benchmarks where it’s now winning, and the market reaction among its Chinese rivals, both suggest this gap is closing faster than a lot of outside observers expected.
