At a time when companies are looking to improve the performance of artificial intelligence while reducing development costs, there is good news for Chinese tech giant Alibaba Group Holding Ltd., which has slashed the cost of training AI search models by nearly 90%. This achievement is due to Alibaba's ZeroSearch technology solution, which is significantly different from other AI search technologies and shows unique innovative advantages.
According to a paper released last week by Alibaba's research team, ZeroSearch, a new technology solution, was officially unveiled. Unlike other AI search technologies, which mostly rely on real search engine interactions to train models, ZeroSearch takes a different approach to improve the search capabilities of AI models through simulation, without the need for actual search engines.
At present, many AI search technologies need to connect query instructions to commercial search engines during the training process, which will incur high costs. The biggest advantage of ZeroSearch is that it can completely avoid such fees. For AI models that have already been trained on a massive knowledge base, ZeroSearch can unleash their potential and allow them to independently generate high-quality query responses. Other technologies may require frequent calls to external search engine data, which is not only costly, but also has problems such as limited data access.
ZeroSearch transforms a reference model into a "search engine-like" that can be used to train other AI systems, dramatically reducing reliance on expensive external search infrastructure. In contrast, some AI search technologies rely heavily on third-party search platforms to provide data support, and the training process will be affected if external cooperation changes or platform rules are adjusted. The self-sufficient training mode of ZeroSearch makes AI search training more "affordable", especially for small R&D teams with limited resources, which is undoubtedly a great boon. For example, sending 64,000 queries to Google through an API costs about $586.70, while using an AI model with 14 billion parameters to generate training responses costs only $70.80, a staggering 88% reduction.
The researchers emphasized that this innovation is of far-reaching significance to promote the development of AI search capabilities, and will effectively promote the improvement of AI model inference and content generation capabilities.
Figure: From simulation training to cost reduction of 88%: Alibaba's ZeroSearch reshapes the AI search landscape
In fact, this research result is highly consistent with Alibaba's strategic layout in the field of AI search. Just a few days ago, Alibaba launched a new and upgraded AI search function for its Quark app. According to Aicpb.com data that tracks the use of AI products, Quark is currently the most popular AI application in China. According to the Quark team, this function, called "deep search", successfully integrates the reasoning ability of Alibaba's self-developed Qwen AI model with online search, and can give accurate answers even if it is a complex user question, and the function has achieved full coverage in multiple terminals such as mobile phones and computers.
It is worth mentioning that the competition in China's AI search field is becoming more and more fierce, and many companies are actively exploring new technologies. For example, Internet search giant Baidu has integrated the R1 inference model developed by start-up DeepSeek into its online search service, providing a summary of search results optimized through model integration. Hangzhou-based DeepSeek and Beijing-based startup Moonshot AI added real-time search capabilities to their chatbots, focusing on instant interactive feedback. Alibaba's ZeroSearch transforms from the source of training, fundamentally solves the problem of cost and independence, and opens up a new path for the development of AI search technology.
In general, Alibaba's ZeroSearch technology solution stands out in the field of AI search technology with its unique training method and cost control advantages, providing new ideas and directions for the development of the industry, and bringing more possibilities for the innovation and popularization of AI search technology in the future.
