China unveiled two new AI models at the World AI Conference (WAIC) 2026 in Shanghai: ScienceOne Omni 2.0, an upgraded science-focused foundation model, and Fenghe, a large open-source model tailored for weather services, public agencies announced on Thursday and Friday.
New AI models at WAIC
The ScienceOne Omni 2.0 was revealed by the Chinese Academy of Sciences (CAS) and aims to strengthen AI systems’ scientific capabilities by removing barriers between general intelligence and specialized expertise. At the same time, the China Meteorological Administration (CMA) introduced Fenghe, and launched a global open-source initiative around the model for weather services.
ScienceOne Omni 2.0: a research-focused foundation model
Developed by a joint team led by CAS, ScienceOne Omni was initially released in July 2025. Version 2.0 introduces several enhancements that enable the model to perform comprehension, reasoning, prediction, and generation across multiple modalities and scientific tasks, according to CAS. The model is described as an “intelligent foundation” intended to provide more reliable and comprehensive support for scientific work.
Fenghe: weather, open source, and international reach
Fenghe, developed by the CMA, is presented as an open-source model dedicated to weather services. It was trained on 50 million tokens drawn from meteorological service data and integrated with so-called “authoritative” meteorological datasets to support forecasting, risk assessment, and weather services. The CMA says Fenghe delivers personalized weather information, service recommendations, and risk alerts for users in China. An international version of the model has also been incorporated into the United Nations’ “Early Warnings for All” initiative to provide weather consultations and risk analyses in Chinese and English.
An open-source ecosystem and very large-scale models
Separately, the startup Moonshot AI unveiled the Kimi K3 model, announced Thursday as the largest open-source model by parameter count (2.8 trillion). According to the company, Kimi K3 supports visual understanding, offers a context window of one million tokens, and has… (original text cuts off here).
