Moonshot AI's Kimi K2.5 Narrows US-China AI Development Gap
- tech360.tv
- 2 hours ago
- 3 min read
Moonshot AI’s newest release, Kimi K2.5, has significantly narrowed the artificial intelligence model development gap between the United States and China. Third-party evaluations indicate this brings the two nations closer than ever in AI capabilities.
This achievement by the Beijing-based start-up raises questions about the effectiveness of US policies, particularly export controls on advanced semiconductors, in restraining China’s AI development efforts. Kyle Chan, a fellow at the Brookings Institution, suggested that for these Chinese start-ups, securing capital appeared to be the primary factor.
Moonshot AI, founded in March 2023, was valued at USD 4.3 billion after raising USD 500 million in its Series C funding round in December. Investors in the organisation include IDG Capital, Tencent Holdings, and Alibaba Group Holding.
Artificial Analysis, a third-party benchmarking firm, reported that Kimi K2.5 scored only a few points behind leading models from US heavyweights OpenAI, Anthropic, and Google DeepMind in comprehensive performance evaluations. Moonshot AI publicly released the weights of its new model, an open-source approach allowing developers globally to download and customise the powerful 595-gigabyte model without subscription costs.
Chan stated that the open-source nature of such a powerful model challenges the strategy of American AI companies that develop closed-source models. He suggested this poses a real challenge to the business models of subscription-based start-ups like OpenAI and Anthropic, given an equally capable, open-source Chinese model.
Kimi K2.5 was also considerably cheaper to operate than its US counterparts, with Artificial Analysis estimating it to be more than four times cheaper than leading models from OpenAI and Anthropic. Moonshot AI indicated that Kimi K2.5 was built on the Mixture-of-Experts architecture, which has aided other Chinese AI companies in achieving efficiency gains over the past year.
US restrictions on access to computational resources have compelled these Chinese companies to maximise the cost efficiency of their models. The new release marked Moonshot AI’s first multimodal model, capable of processing image and video in addition to text.
Artificial Analysis noted that this multimodal capability removes a "critical barrier" to the adoption of open models, as other leading Chinese models have yet to incorporate this feature. A notable characteristic of Kimi K2.5 is "agent swarm," which allows users to manage complex tasks using up to 100 subagents executing workflows in parallel.
Parallel execution capabilities are considered crucial for attracting software engineer users, a lucrative market for AI services. However, the use of many subagents also increases computational costs.
For instance, Chinese company Zhipu AI restricted parallel usage of its flagship coding product due to computational constraints. Moonshot AI has similarly limited the agent swarm feature of its new model to premium users because of concerns regarding available computing power, according to a source familiar with the matter.
Lennart Heim, an AI policy and semiconductor expert, observed that this suggests US export controls have impacted China’s AI development efforts, even if they have not hurt the performance of individual Chinese models. He noted that China has successfully pursued a "fast follower" approach.
Moonshot AI’s Kimi K2.5 model has significantly closed the artificial intelligence development gap between the US and China.
Third-party evaluations indicate Kimi K2.5 performs closely to leading US models while being considerably cheaper to run.
The Beijing-based start-up’s first multimodal, open-source model allows developers to customise its 595-gigabyte architecture without subscription fees.
Source: SCMP
