A KI sign at the MWC Shanghai Tech Show on June 19, 2025.
Bloomberg | Bloomberg | Getty pictures
Beijing – The latest Chinese generative artificial intelligence model that Openai's Chatgpt takes over offers coding functions at a cheaper price.
The startup Moonshot supported by Alibaba was released on Friday evening his Kimi K2 model: an inexpensive, open source model with great language and two factors, which in January undermining Deepseek's industry disorder in China. The open source technology offers free source code access, an approach that only a few US tech giants have chosen as META and Google to a certain extent.
Coincidentally, the CEO of Openaai, Sam Altman, announced in the early Saturday that there would be an indefinite delay in its first open source model again for security reasons. Openai did not immediately answer a CNBC request for comments on Kimi K2.
One of the strengths of Kimi K2 is to write computer code for applications, an area in which companies see the potential to reduce or replace employees with generative AI. Openais US -American Anthropic focused on the coding with its Claude Opus 4 model released at the end of May.
In his publication announcement on the social media platforms X and Github, Moonshot claimed that Kimi K2 exceeded Claude Opus 4 on two benchmarks and achieved a better overall performance than the GPT-4.1 model based on several industry metrics.
“No doubt [Kimi K2 is] A worldwide competitive model, and it is open, “Wei Sun, main analyst for artificial intelligence at counterpoint, said on Monday in an e -mail.
Despite raw improvements in the coding, Sun noticed that Kimi K2 still develops tools for effective integration into other technical systems. This means that it can be more difficult for developers to switch from competing offers to Kimi, which “less than game changing”, she said.
Cheaper option
The new K2 model “also has lower token costs, which makes it attractive for large-scale or budget-sensitive inserts,” said Sun.
The model is available free of charge in contrast to Chatgpt or Claude via Kimis app and browser interface, the monthly subscriptions for your latest AI models.
According to his website, Kimi also calculates 15 cents per 1 million input token and $ 2.50 per 1 million output token. Tokens are a way to measure data for the processing of AI model.
In contrast, Claude opus 4 100 – more for input – 15 USD per million tokens – and 30 – more for the performance – 75 USD per million tokens. In the meantime, GPT-14 calculates $ 2 for every million tokens for input and $ 8 for output.
Moonshot Ai said on Github that developers can use K2 as possible, with the only prerequisite that it displays “Kimi K2” on the user interface if the commercial product or service has more than 100 million active users or corresponds to $ 20 million per month.
Hot, you have a market
The first reviews of K2 in both English and Chinese social media were largely positive, although there are some reports of hallucinations, a prevailing problem in the generative AI, in which the models make up information.
Nevertheless, K2 is “the first model that I use in production since Claude 3.5 Sonnet,” said Pietro Schirano, founder of Startup Magicpath, who offers AI tools for design, in a contribution to X.
Moonshot has some of his former AI models open. At the beginning of last year, the company's chat bot was raised as an alternative China for Chatgpt, which is not officially available in the country. But similar chatbots from Bytedance and Tencent have been launched since then, while the Tech -Riese Baidu has revised its core search engine with AI tools.
Kimi's latest AI publication is made because investors look Chinese alternatives to US Tech in the global AI competition.
Despite the excitement via Deepseek, the privately owned company must have announced a large upgrade of its R1 and V3 model. In the meantime, Manus Ai, a Chinese startup that was created at the beginning of this year as a further start of the Type Deepseek, moved his headquarters to Singapore.
Openai GPT-5 does not yet have to reveal in the USA.
Work on GPT-5 may take on technical resources and prevent Openai from progressing on his open source model.
Grok 4 competitors
Kimi K2 is not the only recent publication of the company. Moonshot launched a Kimi research model last month and claimed that it matched the 26.9 point number of Google Gemini Deep Research and defeated the version of Openai on a benchmark called “Humanity's Last Exam”.
The Kimi research model was even mentioned last week during the XAI publication of Grok 4 in Elon Musk, which achieved 25.4 for himself at the Benchmark “Humanity's Last Exam”, but achieved a 44.4 point number if they were allowed to use a variety of KI tools and web search.
“Kimi-Resarcher represents a paradigm shift in the agent AI,” said Winston Ma, extraordinary professor at the NYU School of Law. He referred to AI's ability to make several decisions at the same time to do a complex task.
“Instead of only creating flowing reactions, it shows autonomous thinking at expert level – the kind of complex cognitive work that previously lacked LLMS,” said Ma. He is also the author of “The Digital: How China's Technical Power shapes the future of AI, Blockchain and Cyberspace”.
– Victoria Yeo from CNBC contributed to this report.