China introduces upgraded AI foundation model to accelerate scientific discovery

China unveiled on Friday an upgraded AI foundation model designed to accelerate scientific discovery across disciplines at the ongoing World Artificial Intelligence Conference in Shanghai, marking the country's latest push to integrate artificial intelligence into fundamental research.

ScienceOne Omni, developed by the Chinese Academy of Sciences with contributions from multiple institutes, is built as a specialized intelligence infrastructure for scientific tasks, aiming to break down barriers between general artificial intelligence and domain-specific scientific capabilities.

Building on the original ScienceOne foundation model launched at the conference in July 2025, the new version represents a significant upgrade by fusing broad AI capabilities with deep scientific expertise in a unified architecture that enables the model to work across different scientific modalities and tasks.

At its core, a progressive three-layer structure — scientific data encoding, natural world knowledge alignment, and domain-specific task decoding — enables a single model to handle cross-disciplinary data understanding, knowledge reasoning, scientific prediction, and content generation.

Trained on 8 million high-quality scientific reasoning samples covering more than 200 research tasks, the model encoded scientists' reasoning patterns into verifiable inference chains.

Evaluations across more than 60 professional research tasks showed that ScienceOne Omni significantly outperformed flagship models such as Gemini-3.1-Pro and GPT-5.5 on most benchmarks, while achieving state-of-the-art results on domain-specific tasks, including chemical property prediction, spectral-to-molecular structure prediction, and protein site forecasting.

An integrated intelligent research platform has been built around the model, pooling scientific datasets with more than 8,000 professional research tools. A literature review agent can generate professional-level reviews within three hours with 90 percent evidence attribution accuracy, having already produced nearly 40,000 reviews. The entire computing architecture has been adapted to self-reliant domestic chips, including Ascend and Hygon.

Demonstrations at the launch showcased its applications in different fields. In mechanics, an AI-powered fluid simulation tool delivers results within 10 seconds with less than 5 percent error.

In astronomy, a spectral agent deployed at the National Astronomical Observatories improved rare celestial object identification by around 50 percent and now supports the LAMOST telescope — home to the world's largest stellar spectral dataset.

In chemistry, the model drives a "machine scientist" system that autonomously completes the full loop from literature mining to experimental synthesis, having screened millions of virtual candidates with eight of the top 20 outperforming the previous best material.

ScienceOne Omni has been adopted by more than 50 CAS' institutes, as well as more than 30 universities, research institutions and State-owned enterprises, and is expanding internationally through UNESCO and the Alliance of International Science Organizations.