Alibaba-backed Moonshot releases Kimi K2 AI rivaling ChatGPT, Claude
On June 19, 2025, a sign of AI at the MWC Shanghai technology show.
Bloomberg | Bloomberg | Getty Images
Beijing – The latest Chinese productive artificial intelligence model to undertake Openai’s Chatgpt offers coding features at a lower price.
Alibaba supported start moonshot Released late on Friday night Some K2 Model: A low-cost, open-source large language model-two factors that support the industrial deduction of China-based Deepseek in January. Open -source technology provides free resource code access, which is an approach from a small number of US technology giants other than Google, to some extent.
Coincidentally, Openai CEO Sam Altman Announced early on Saturday The first open source model will be an indefinite delay due to security concerns. Openai did not respond immediately to a CNBC request for a comment on some K2.
One of the powerful aspects of some K2 is writing computer code for applications that the personnel of enterprises see the potential to reduce or replace the productive AI. Openai’s US rival Anthropic Claude Opus 4 focused on coding with the model It was released in late May.
In the broadcasting announcement in social media platforms in X and Github, Moonshot claimed that some K2 has overcome the Claude Opus 4 in two criteria, and Openai has a better general performance than the coding GPT-4.1.
“No doubt [Kimi K2 is] Globally a competitive model and open source, “the contradictory wei Sun, the chief analyst of artificial intelligence, on Monday in an e -mail.
Cheaper option
“Beyond that, there are lower coin costs, which makes it attractive for large -scale or budget -sensitive distributions,” he said.
Unlike Chatgpt or Claude, which charges monthly subscriptions for the latest AI models, the new K2 model can be used through Kimi’s application and browser interface.
Some, according to the website, only 15 cents and 1 million output for each 1 million input coin cost $ 2.50 per coin. The coins are a way Measurement of data for artificial intelligence model processing.
In contrast, Claude Opus 4 100 times more fees For input – $ 15 per million to coin – and 30 times more for output – $ 75 per million. By the way, for each million coins, GPT-4.1 Fees $ 2 for input and $ 8 for output.
MOONSHOT AI He said in Github Developers can use K2 as they wish, commercial products or services have more than 100 million active users per month, or if they make a $ 20 million equivalent in monthly income, they can use “K2” in the user interface.
Hot AI market
Although K2’s initial investigations about both English and Chinese social media are largely positive Some hallucination reportsIt is a common problem in the productive AI, where models form information.
Nevertheless, K2, “Claude 3.5 Sonnet since the first model I feel comfortable in production” Pietro Schirano, Founder of Startup Magnique In an article on X, AI offers AI tools for design.
Moonshot provided some of the previous AI models. The company’s chat boat gained popularity in early last year as an alternative to Chatgpt, which is not officially found in China. However, similar chat boots from ByTEDANCE and TENCENT made the market crowded and the technology giant Baidu renewed the core search engine with AI vehicles.
Kim’s latest AI version comes to investors’ US technology in the global AI competition, taking into consideration the Chinese alternatives.
Nevertheless, despite the excitement of Deepseek, the private company has not yet announced a major upgrade to the R1 and V3 model. Meanwhile, Manus AI, a Chinese initiative that emerged as another deepseek type earlier this year He placed his center in Singapore.
Openai de GPT-5 in the US has not yet put forward.
Studies on the GPT-5 may be receiving engineering resources by preventing the progression of Openai’s open-source model, the Sun of the Convention, adding that it is difficult to publish a powerful open source model without weakening the competitive advantage of a registered model.
Gok 4 contestants
Some K2 is not the latest version of the company. Moonshot launched some research model last month and claimed that Google’s Gemini Deep Research was matched with 26.9 points. Openai’s version was defeated on a criterion called “The Last Examination of Humanity”.
Some research model took a word during Elon Musk’s last week XAI version of Grok 4 – He received 25.4 points on his own in the comparison of the “final exam of humanity”, but He obtained 44.4 points when he was allowed to use various AI vehicles and web search.
“Some-Research represents a paradigm change in the Agent Artificial Intelligence,” NYU Law Faculty Professor Winston Ma said. He said. He was talking about AI’s ability to make a few decisions on his own to complete a complex task.
“Instead of producing fluent reactions only, autonomous reasoning at the expert level – the complex cognitive work that has previously missed LLMS.” He said. It is also the author of “Digital War: How China’s technology power shapes the future of AI, Blockchain and cyber.”
– Victoria Yeo from CNBC contributed to this report.
