Alibaba Group Holding says it has found a way to shrink the costs of training search‑focused artificial intelligence models by almost 90 per cent.
In a paper posted on the arXiv server last week, researchers at the Chinese technology company describe “ZeroSearch”, a training shortcut that lets large language models improve their search skills at a fraction of the cost.
Alibaba argues that the approach removes one of the most expensive steps in reinforcement learning: routing tens of thousands of queries through paid application programming interfaces.
ZeroSearch puts such bottlenecks to use. A reference model that already holds a broad knowledge base acts as a stand‑in search engine, producing both helpful and deliberately noisy documents for each query. Newer models are then trained on those self‑made noisy documents.
The fine‑tuning starts with plainly worded prompts and clean answers. Over time, a “curriculum rollout” adds clutter, typos, and contradictions so that the trainee model experiences something closer to the open internet.
According to the paper, the data noise strengthens reasoning and teaches the system to survive in messy data environments. “Our key insight is that LLMs have acquired extensive world knowledge during large‑scale pretraining and are capable of generating relevant documents given a search query,” the researchers explain.
Teams do not have to pay external providers for every query. The change is especially significant for smaller teams that have limited resources. For instance, sending 64,000 queries to Google through an API can cost about US$586.70. Running the same training with a 14‑billion‑parameter AI model cost roughly US$70.80, an 88 per cent saving.
Alibaba’s ZeroSearch does not require additional hardware
ZeroSearch, the team argues, does not add extra hardware demands. It relies on standard supervised fine‑tuning, so cloud compute needs stay within the range of typical development budgets.
The paper has already drawn attention inside academic machine learning circles. Industry observers say the savings could influence upcoming budget cycles during the coming financial year.
Industry rivals have been racing to trim the costs of large language model training as they push new generative AI products.
Alibaba’s new feature arrives as Chinese firms compete with US firms such as Google and OpenAI, both of which have invested heavily in AI agents. By reporting a near-tenfold cost cut, Alibaba signals that the cost for AI search might soon shift.
Cryptopolitan Academy: Want to grow your money in 2025? Learn how to do it with DeFi in our upcoming webclass. Save Your Spot