Breaking news! Surpassing Google search engine! Alibaba open-sources search engine model ZeroSearch! 2025
Alibaba's open source search engine ZeroSearch opens a new era of AI information retrieval!
Core content:
1. ZeroSearch uses large model pre-training knowledge to achieve self-search capabilities
2. No external search engine is required, reducing costs by 87.93%, and performance exceeds Google
3. From "data transfer" to "cognitive modeling", deep integration of knowledge and retrieval is achieved
When large models learn to "self-search": ZeroSearch will reconstruct the information retrieval paradigm in the AI era
In 2025, when the competition for big model computing power is becoming increasingly fierce, Alibaba Tongyi Laboratory's open source framework ZeroSearch is like a huge rock thrown into a lake, causing a huge stir in the field of AI search.
ZeroSearch mainly utilizes the rich knowledge accumulated by large models during large-scale pre-training and transforms it into a retrieval module that can generate relevant content based on search queries. At the same time, it can also dynamically control the quality of generated content, which is a special function that traditional search engines do not have.
The researchers conducted a comprehensive evaluation on seven major question-answering datasets, including NQ, TriviaQA, PopQA, and HotpotQA. The results showed that after using ZeroSearch, a supervised fine-tuning model with 7 billion parameters had a search capability of 33.06; a model with 14 billion parameters reached 33.97, exceeding Google Search's 32.47.
This reinforcement learning framework, which does not rely on a real search engine, not only surpasses Google (32.47) with a search capability score of 33.97, but also quietly starts a technological revolution from "external call" to "autonomous construction" by deeply exploring the intrinsic retrieval capabilities of large models.
Its core value lies in converting the knowledge graph accumulated in the pre-training stage of large models into a dynamically controllable "digital search engine", allowing the AI system to have the ability to complete the "search-inference-verification" closed loop within its own knowledge system without external APIs for the first time.
1. Breaking the ice with technology: a paradigm revolution from “data handling” to “cognitive modeling”
Traditional AI search relies on the "retrieval augmented generation" (RAG) architecture, which requires real-time calls to external search engines such as Google and Baidu to obtain documents.
This model has three limitations:
The breakthrough of ZeroSearch lies in the construction of an "endogenous retrieval engine": using the world knowledge (such as entity relationships and factual data) learned by the large model in pre-training, it generates simulated documents through lightweight supervised fine-tuning (only thousands of labeled samples are required).
These documents are not simple random noise, but the content quality is precisely controlled through prompt engineering: in the early stages of training, the model generates highly relevant "golden documents" to help understand the basic retrieval logic. As the training progresses, "interference documents" (such as confusing information containing similar entities) are gradually added, forcing the model to improve its semantic discrimination capabilities.
This “curriculum learning” mechanism follows the cognitive law from easy to difficult, just like how humans transition from dictionary retrieval to complex literature research, allowing the model to achieve an exponential improvement in search capabilities within 64 training cycles.
In terms of technical implementation, ZeroSearch designed a three-stage interaction template:
Reasoning stage (thinking): The model first generates a preliminary reasoning path based on the query, such as "to answer 'the year when the theory of relativity was proposed', we must first locate Einstein's academic timeline." Simulated search (Search): Generate 3-5 simulated documents based on the reasoning path, including the correct answer and interference items (such as mixing 1905 and 1915 (when general relativity was proposed)). Verification generation (answer): By comparing the document content with internal knowledge, the final answer is output and the confidence level is annotated.
This structured interaction not only improves training efficiency (four A100 GPUs can support a 14 billion parameter model), but also allows the model to learn to "think about search strategies like humans" - when processing "which country is the Mona Lisa currently in", it will prioritize retrieving the "location of the Louvre" rather than directly calling up memory. This accumulation of procedural knowledge is exactly what the traditional RAG architecture cannot achieve.
2. Cost disruption: Industry restructuring from “money-burning competition” to “efficiency revolution”
Another disruptive aspect of ZeroSearch is the reconstruction of training costs. In traditional methods, every time a model learns a retrieval strategy, it has to pay a "data tax" to the real search engine. However, ZeroSearch reduces this cost to less than 1/5 of the original cost through "knowledge internalization":
Experiments show that at the same training scale, the cost of using ZeroSearch is only US$70.80, which is 88% lower than SerpAPI's US$586.70.
The chain reaction brought about by this cost advantage is rewriting the industrial ecology of AI search:
Technological equality for small and medium-sized institutions: Search optimization projects that previously required a million-dollar budget to launch can now be reduced to a few thousand yuan on Alibaba Cloud GPU cloud servers, driving the explosion of AI applications in vertical fields (such as legal retrieval and medical literature analysis). The return of data sovereignty: Enterprises do not need to disclose query logs to third parties and can fully control training data in a private cloud environment, which is crucial for sensitive fields such as finance and government affairs. A bank's actual test shows that the internal risk control information retrieval system based on ZeroSearch has increased response speed by 40% and reduced data leakage risk by 65%. Shift in R&D paradigm: The traditional “model training + external API call” fragmented model is being replaced by the closed loop of “pre-training knowledge activation + simulated search enhancement”. Tongyi Lab data shows that this framework shortens the model iteration cycle from 45 days to 12 days, making “weekly version updates” possible.
3. Ecological Game: The “Three-Body War” Behind the Open Source Strategy
The open source of ZeroSearch is essentially Alibaba’s strategic layout in the “Three-Body Game” of AI search:
Technical dimension: Dimensionality reduction attack of hybrid architecture Faced with the closed-source barrier of Google's "index library + commercial API", the "large model endogenous search" represented by ZeroSearch has opened up a new track.
While open source models such as DeepSeek-R1 achieve autonomous optimization of search logic through reinforcement learning, the traditional model that relies on trillions of web page indexes is showing signs of fatigue - the latter requires continued investment of billions of dollars to maintain the crawler system, while the former only needs to inject high-quality knowledge bases (such as Wikipedia structured data) into pre-training data to dynamically generate training samples through simulated search.
This "light asset, heavy intelligence" model is just like the disruption of fuel vehicles by new energy vehicles, allowing latecomers to directly enter the intelligent fast lane without having to build a huge "information gas station" network.
Business dimension: From traffic harvesting to value symbiosis, Alibaba’s “two-front battle” strategy is clearly visible:
Regulatory dimension: building a bridge between innovation and compliance
As the EU AI Act requires that “generative AI needs to disclose the source of training data”, ZeroSearch’s “data self-generation” has become a compliance advantage - its simulation documents are all generated based on public knowledge bases, avoiding the privacy data that may exist in real search engines (such as user search history), clearing obstacles for global market access. The CIO of a multinational company pointed out: “Under the strict constraints of GDPR, ZeroSearch has improved the compliance of our cross-border search system by an order of magnitude.”
4. Challenges and the future: When search capabilities evolve into “digital instincts”
However, there are still unresolved issues behind the technological breakthroughs:
The paradox of knowledge boundaries: The pre-trained knowledge of large models ends in 2023, and the simulation documents generated by ZeroSearch cannot contain real-time data (such as new scientific research results released in 2025). Alibaba's solution is a "dual-mode architecture" - daily problems are handled by endogenous search, and real-time needs are seamlessly switched to external APIs. This "dynamic and static combination" model is being verified in the internal test of Quark Search. Quality control accuracy: Although the F1 score reward mechanism improves the accuracy of answers, in extreme noise scenarios (such as deliberately confusing "Tesla CEO" with "SpaceX CEO"), the model still needs more reasoning steps to correct errors. Tongyi Lab is developing a "multi-evidence cross-validation" module to perform logical verification by introducing internal knowledge graphs. Challenges of ecological collaboration: The open source framework requires the community to contribute high-quality prompt templates and industry data sets. Whether Alibaba can replicate the success of Linux depends on whether it can establish an effective developer incentive mechanism. Its current "Search Innovation Plan" has provided tens of millions of computing power subsidies and attracted more than 300 teams to join in the first batch.
5. Conclusion : Redefining the essence of “search”
The real value of ZeroSearch lies in revealing the essential transformation of "search" in the era of big models : it is no longer a simple information capture, but an organic integration of the dynamic call of the knowledge system and the reasoning process.
When AI can simulate search scenarios in its own "brain" , it is equivalent to humans mastering the ability to "traverse the library with eyes closed and meditating". This evolution of cognitive mode is reshaping the interactive interface between people and information .
From the perspective of technological history, this may be the most revolutionary breakthrough in the search field after Google's PageRank algorithm . The former solves the problem of "how to retrieve information from the Internet" and the latter solves the problem of "how to let AI independently build retrieval capabilities."
As the number of stars of the ZeroSearch code on GitHub exceeds 10k, this technical experiment that started in Alibaba Labs is evolving into a paradigm revolution in which A developers around the world participate.
When search capabilities become the "digital instinct" of large models, we may be ushering in an intelligent era in which the efficiency of information acquisition is exponentially improved, and this is just the beginning.
