metamorworks - stock.adobe.com
China startup Moonshot AI rivals U.S. with cheap open model
A Chinese startup is out with an open source trillion-parameter model that excels at coding and agentic tasks. It outperforms Western competitors on some benchmarks.
As geopolitical and economic tensions grow between the U.S. and China, another Chinese AI vendor released a new large language model that offers the same capabilities, apparently at a lower price than OpenAI's LLMs.
Moonshot AI, a startup backed by Chinese tech giant Alibaba, released the Kimi K2 model, an open source LLM, on July 11.
Kimi K2
Kimi K2 has 32 billion activated parameters and one trillion total parameters, according to Moonshot AI. It comes in two versions: Kimi-K2-Base and Kimi-K2-Instruct. The startup said the base model suits researchers who want complete control for fine-tuning and custom applications.
Meanwhile, the Instruct version is a post-trained model for general-purpose chat and agentic applications. The startup cited benchmarks that show that the Kimi models perform better than DeepSeek-v3-0324 from another Chinese startup and Open AI’s GPT 4.1 in agentic and competitive coding on some benchmarks, but still lag behind Claude 4 Opus on others.
Kimi K2's release evokes similar reactions as DeepSeek's first release of its R1 model earlier in the year. Moonshot released Kimi K1.5 days after DeepSeek released DeepSeek-R1 in January. The vendor’s open source models provide an alternative to proprietary models like those from OpenAI and raise questions about the inexpensiveness of AI technology. It also comes days after another Chinese vendor, Baidu, made its Ernie 4.5 model open source.
"Moonshot's choice of open licensing, affordable API tiers, and optional self-hosting positions them to attract a global developer base and enterprise user ecosystem, which is much needed to counter proprietary Western models," said Arun Chandrasekaran, an analyst with Gartner.
He added that the model's mixture of expert architecture and chip innovations can connect lab research and real-world efficiency. For example, Moonshot provided an example of Kimi K2 being used to explore remote-work salaries.
Moreover, Kimi K2 is priced below OpenAI comparable models at the non-cached price $0.60 per million input tokens and $2.50 per million output tokens. Meanwhile, OpenAI’s pricing is $2 per million input tokens and $8 per million output tokens.
Challenges and skepticism
However, the model's performance does not mean it's without challenges.
"As a Chinese-developed model, overseas adoption may face regulatory or perception challenges amid shifting global tech dynamics," Chandrasekaran said.
Moreover, Moonshot's K2 model might be an attempt to undermine the U.S. model market, said David Nicholson, an analyst with Futurum Group.
"When companies outside of the U.S. sell products into our markets for less than the cost to produce those products, we call it dumping," Nicholson said. "It's a strategy to dominate a particular market. Flood the market with ‘open’ and ‘cheap.’”
He added that the big question is the actual cost of models like Kimi K2.
"China and the U.S. are in a trade battle," Nicholson continued. "Announcements about massive cost differences should be met with skepticism."
Enterprises should be careful about surrendering their data to China,” he said.
Meanwhile, OpenAI revealed on July 12 that it will delay the release of its open source model indefinitely. CEO Sam Altman said the vendor needs to run more safety tests.
And Meta, known for its open source model, Llama, continued its goal of forging ahead in AI technology. The company recently acquired a voice generative AI vendor called Play AI. Play AI is known for producing realistic, natural-sounding voices with AI technology.
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
