The AI ​​world is really lively in 2025. Not long after OpenAI released Sora, Alibaba made another move - officially open-sourcing ZeroSearch, an AI system that can self-learn search , which directly overturned the idea of ​​"training AI with search engines".

What does it mean? In the past, training large models required searching APIs like Google. If you asked one question and one API, you would go bankrupt if you asked too many questions. Now, Alibaba has simply taught AI to "be its own search engine". Training no longer relies on external calls, and even Google's money has been saved.

This is not about reducing costs and increasing efficiency, but about changing the mindset, modifying the foundation, and redefining search.

Who needs Google when big models have built-in search?

The traditional approach is to let AI search before answering questions, obtain some web pages and documents, and then understand and generate. This is the so-called RAG (retrieval-augmented generation) model. It sounds advanced, but it is actually very dependent on external interfaces. It has to work for whoever has the strongest search engine.

ZeroSearch directly abandons this approach and embeds the "retrieval" skill into the large model itself. In other words, this thing can "fill in" search results by itself without having to go online to call external searches.

Let the data speak for itself: Previously, one training run required 64,000 Google searches, which cost nearly $600. Now, with ZeroSearch, the cost has been reduced to just over $70. This is nearly 80% cheaper!

What’s even more amazing is that its performance is even better than Google Search itself. In multiple standard question-answering tests, Alibaba’s 14 billion parameter model retrieval scores simply surpassed Google’s.

It’s not a small optimization, it’s a low-level reconstruction

ZeroSearch can achieve this not by simply “saving money”, but by rebuilding the entire technical architecture.

It uses a "hybrid reasoning architecture" with three core features:

1. Built-in search capabilities

No longer outsourced, the search module is directly embedded in the model. It can also "simulate" noisy documents and dynamically adjust keyword prompts, so that the results are more relevant and reliable. For example, cross-border e-commerce merchants can use it to generate product descriptions with industry tone, and even user reviews can be automatically completed.

2. The learning path is like taking a class

The model does not start with the most complex tasks, but starts with simple problems and gradually challenges more difficult tasks. This "curriculum learning" method combined with reinforcement learning (PPO algorithm) allows a small model with 7 billion parameters to match Google and directly surpass the 14 billion parameter model.

3. Multimodal and real-time

It can search for pictures, text, voice, and video, and also supports streaming processing. For example, if a primary school student asks a question using voice, it can generate an answer with pictures and videos in real time. It makes it easier for parents to help with homework.

It’s not a technology show, it’s an industry brand

Alibaba is not just doing a demo to show off its skills, but really wants to "decentralize" this system to everyone who can use it.

Let’s talk about small and medium-sized enterprises first.

Think about it, in the past, every time you call a search API, you have to pay a fee, and the company will lose hundreds of thousands of dollars a year. Now with ZeroSearch, the model searches by itself, and you don’t have to spend that money. Cross-border e-commerce uses it to build a recommendation system, which directly saves half of the budget a year.

Then there is the vertical scene.

For example, in e-commerce, after Alibaba International Station connected to ZeroSearch, the efficiency of product matching increased by 40%. It is also very popular in the field of education. It can generate interactive teaching materials and 3D experimental animations, and improve students' understanding efficiency by 25%. After using it, the internal document search of enterprises has been reduced from 5 minutes to 30 seconds.

Another example is AI companies.

In the past, training question-answering systems required preparing a large amount of labeled data. ZeroSearch allows you to prepare 60% less data and improve the accuracy by 10%.

This is the triple combo of “cost reduction + efficiency improvement + liberation of productivity”.

The real power of open source lies in the ecosystem

ZeroSearch is not a closed-door project; it is open source, which means that developers around the world can participate.

Three days after its launch, the number of GitHub stars exceeded 10,000, and Hugging Face was downloaded more than 500,000 times, with the community reaching its peak. A developer directly contributed a "dynamic quality control plug-in", which immediately increased the relevance of model-generated content by nearly 20%.

This kind of open source ecosystem collaboration is the basis for Alibaba to widen the gap.

What's even more amazing is that Alibaba has also formed a joint effort with NVIDIA - a two-pronged approach of algorithm + computing power. NVIDIA used Tongyi Qianwen as the inference model, and Alibaba used ZeroSearch to make up for the search shortcomings. After the cooperation between the two sides, the memory usage was reduced by 40%, and the inference speed was increased by a quarter.

Technology is not something that can be achieved alone. Only with cooperation, feedback and an ecosystem can one go far.

The challenges ahead are not small

Of course, this road is not all smooth sailing.

For example, ZeroSearch is not stable enough to handle long texts (more than 1,000 words), and the accuracy rate will drop by more than 10%. The response delay of complex multimodal retrieval (such as image + audio + foreign language) is still a bit high, and there is still room for improvement before it can be truly "real-time".

Looking at the competitors, Tencent's HunyuanCustom once surpassed ZeroSearch by 15% in video generation, and Huawei Ascend also has a deep layout in AI search.

Another issue is law and compliance. Who owns the content generated by the big model? What if false information appears? All of this requires a system to keep up. Fortunately, Alibaba has jointly released a technical governance report with the China Electronics Technology Standardization Institute, and data tracing and content review are also being done.

The real change is paradigm-changing

If AI used to be fed by data and backed by search engines, then ZeroSearch represents the enlightenment of AI's "independent thinking" ability.

In the future, it will not only be able to run locally on mobile phones and IoT devices, but also enter AR/VR scenarios and even undertake AI ethics checks. Alibaba plans to introduce a "value alignment" module to ensure that the content it generates is close to our mainstream values.

From technology to industry and then to governance, this move goes far enough.

ZeroSearch is not just a card played by Alibaba, but more like a major rewrite of the entire AI development paradigm. AI no longer relies on search, but instead masters search capabilities on its own. It's like teaching a child to learn to walk on his own, rather than holding him in his arms for life.

At this moment, Alibaba is not a follower but a leader.

And the story of ZeroSearch has just begun.