Alibaba Cuts Search AI Training Cost by 88% with ZeroSearch Technology

Chinese tech giant Alibaba Group Holding has announced a breakthrough in reducing the cost of training artificial intelligence (AI) models for search by nearly 90%. The achievement will allow companies to expand AI capabilities while reducing development costs. Alibaba researchers have introduced a new approach called ZeroSearch, which improves AI search capabilities through simulations, eliminating the need to interact with real search engines. This method eliminates the high costs associated with using commercial search engines, allowing AI models already trained on big data to generate high-quality answers to queries.

ZeroSearch transforms a reference model into a search engine that can train other AI systems, reducing the reliance on expensive external infrastructure. This makes training search models more affordable, especially for smaller teams. For example, sending 64,000 queries through Google’s API costs about $586.70, whereas using an AI model with 14 billion parameters reduces the cost to $70.80 — a savings of 88%.

The new technology could greatly enhance AI-powered search capabilities, which the researchers say is critical to improving reasoning and answer generation. Alibaba is already making progress in this area: An update to Quark, China’s most popular AI app, now includes a “deep search” feature that combines Qwen’s reasoning model with online search to accurately answer complex queries.

Other Chinese companies are also pushing forward with AI search. Baidu has integrated DeepSeek’s R1 model for aggregated results, while startups DeepSeek and Moonshot AI have added real-time search capabilities to their chatbots to provide better answers.