Chinese artificial intelligence (AI) startup MoonshotAI's recently unveiled large language model (LLM) "Kimi K3" has emerged as a new variable in the global AI market. Amid assessments that the domination of U.S. AI corporations is wavering, the phrase "Kimi shock" has appeared following the "DeepSeek shock." MoonshotAI had little presence early this year as it was overshadowed by DeepSeek, but analysts say a sweeping overhaul of its business strategy to respond quickly to the AI competitive landscape changes sparked by DeepSeek led to Kimi's popularity.
According to Chinese business outlet Caixin on the 21st, Kimi K3, immediately after its unveiling on the 17th, was recognized by global AI performance evaluator Artificial Analysis as having performance close to the latest models from OpenAI and Anthropic. At the World Artificial Intelligence Conference (WAIC) in Shanghai, which wrapped up the day before, more visitors flocked to MoonshotAI's booth than to Alibaba's or Tencent's, and the company temporarily suspended new paid subscriptions as users surged.
Bloomberg said this was "the biggest AI buzz since DeepSeek," and analyzed that Kimi K3's success rests on three strategies. They are: ▲ a pivot to focus on developers and corporate clients ▲ a shift to open weights ▲ development of ultra-large models.
◇ Turning the "DeepSeek shock" crisis into an opportunity… a bold move to target developers
After first launching Kimi in 2023, MoonshotAI was once counted among the leading Chinese AI startups. But the landscape changed abruptly when DeepSeek announced "R1" on the day MoonshotAI unveiled its "K1.5" early this year, monopolizing market attention. While Kimi K1.5 is a closed model used internally, DeepSeek R1 is open weight, disclosing the model's core weights so developers can use them freely.
With this open-weight strategy and its low cost and high performance, DeepSeek R1 drew praise from developers at home and abroad, and Chinese big tech such as Tencent quickly adopted DeepSeek's models in their own services. During the same period, Kimi had to endure user churn.
In response, MoonshotAI founder Yang Zhilin overhauled the strategy. Instead of chatbot competition aimed at general consumers, the company shifted to a focus on developers and corporate clients. At the same time, it decided to release its next flagship model with open weights. Based on this strategy, Moonshot released K2 in July last year, and quickly secured corporate clients as U.S. AI services such as Cursor and Perplexity adopted K2. K3 also maintains the open-weight strategy and is evaluated to have significantly improved coding and AI agent performance.
Another bold move is the ultra-large model. While Chinese AI corporations, hampered by U.S. semiconductor export controls, focused on small models due to difficulties securing graphics processing units (GPUs), MoonshotAI instead scaled up. It expanded a model initially reviewed at around 200 billion parameters to 1 trillion parameters, and unveiled a 2.8 trillion-parameter model with K3. Yang is said to believe that, in the long term, larger model size will confer greater competitiveness in tasks such as reasoning, coding, and AI agents.
◇ The ripple effect of the "Kimi shock"… Chinese corporations reassessed, U.S. sees price competition
The impact of K3 has spread across China's AI industry. According to Caixin, immediately after K3's unveiling, competitors Zhipu and MiniMax saw their share prices plunge. Zhipu fell more than 28% on the 17th alone, the biggest drop since its listing, and MiniMax also fell more than 15% the same day. Huachuang Securities analyzed, "As K3 raised the performance ceiling for open-weight models, the market has begun to reassess the value of AI model corporations."
Developers' responses are also hot. On AI platform OpenRouter, K3 climbed into the upper ranks for token usage within three days of its release, surpassing some of Anthropic's flagship models. After K3's release, the company temporarily suspended new paid subscriptions to prioritize compute resources for existing users as requests vastly exceeded expectations.
The "Kimi shock" is also likely to put considerable pressure on U.S. AI corporations. In a report on the 17th, global investment bank (IB) Bernstein said K3 is priced at $3 per 1 million input tokens and $15 per 1 million output tokens, up to 70% cheaper than some of Anthropic's models. Bernstein noted, "The narrowing gap in inference performance among state-of-the-art AI models is a negative signal for corporations' profitability," adding, "OpenAI and Anthropic have also, in recent weeks, entered competition on pricing and usage limits."
Riding its steep growth, MoonshotAI is seeking a large fundraising. In Feb., it secured investments in quick succession, pushing its valuation to $20 billion (about 29 trillion won), and it is now pursuing new funding targeting a valuation above $30 billion (about 44 trillion won). It is said to be aiming to list on the Hong Kong stock market within as soon as six months.