Alibaba has developed a new training method for large language models (LLMs) that makes search functions in AI systems way cheaper to train. The ZEROSEARCH solution replaces costly API queries to external search engines with a simulation-based process, cutting training costs by almost 90 percent. At the same time, the performance of the models remains at a high level.

With ZEROSEARCH, Alibaba is responding to a key challenge in training modern AI: traditional reinforcement learning methods for search tasks rely on millions of search engine queries. These not only generate high fees, but also deliver results of varying quality – a problem for the efficiency and scalability of training.

Simulation instead of live search

The method developed by Alibaba consists of two steps: First, a language model is refined through supervised fine-tuning to create a retrieval-capable system. It can generate relevant documents in response to queries, simulating the behaviour of real search engines.

In the subsequent reinforcement learning phase, the level of difficulty is gradually increased through a curriculum-based rollout concept: the quality of the simulated documents decreases in a controlled manner, forcing the model to continuously improve its relevance assessment.

The results show that ZEROSEARCH is not only cost-efficient, but also powerful: a Qwen2.5-7B model trained with ZEROSEARCH achieved retrieval performance on par with Google Search. The larger 14B version even outperformed it – with 88 percent lower training costs.

Goal: Access to powerful AI for more companies

‘With ZEROSEARCH, we are drastically reducing the cost of training large language models to simulate search engine behaviour. This enables developers, and small and medium-sized enterprises in particular, to build their own reinforcement learning frameworks without the need for expensive queries to external search engines,’ says Huang Fei, head of the Tongyi Natural Language Processing Lab at Alibaba. ‘ZEROSEARCH is an important milestone in the democratisation of large-scale reinforcement learning technologies: powerful, efficient and significantly more affordable.’

Alibaba is not only pursuing this approach internally. The company has already made a large number of its own AI models available as open source – in various languages, sizes and application areas. Alibaba is thus explicitly targeting the international developer community.

Top marks in independent analysis

According to an analysis by Artificial Analysis, the Qwen3-235B-A22B model, the latest addition to the Qwen series, ranks first in the cost category and fifth in the intelligence category (computing, programming, logic, natural sciences). This positions Alibaba among the leading providers worldwide.