China National Petroleum’s Kunlun Large Model Unveiled: A 300-Billion Parameter AI Collaboration with Huawei, China Mobile, and iFLYTEK

In a significant move that underscores China’s burgeoning AI capabilities, China National Petroleum Corporation (CNPC) has unveiled its 300-billion-parameter Kunlun large language model. This marks a pivotal step for the energy giant into the realm of artificial intelligence, promising to reshape operations and unlock new efficiencies across its vast enterprise.

Developed in collaboration with partners including China Mobile, and a team comprising some of the leading tech firms. The Kunlun model completed its registration in August 2024, making it the first of its kind approved for use in China’s energy and chemical sectors. This launch builds on the foundation laid by prior Kunlun model iterations, including a 33-billion-parameter model introduced last August and a 70-billion-parameter version released in November.

The latest 300-billion-parameter Kunlun model boasts not only a significant leap in scale but also the simultaneous deployment of 100 application scenarios. These applications span a range of operations, with 43 receiving updates and 57 being newly introduced, all of which are now fully operational.

Furthermore, the collaboration extends to the development of specialized AI solutions. Working closely with Huawei, CNPC has established a visual AI model tailored specifically for the oil and gas industry. This initiative involved the labeling and training of over 700,000 industry-specific visual datasets, with the model’s parameters reaching 4.4 billion. This visual AI model, leveraging Ascend hardware and Huawei Cloud Stack hybrid cloud technology, aims to enhance its industrial vision capabilities and delve deeper into high-value applications.

Supporting the deployment of this advanced AI infrastructure is the Kunlun AI Middle Platform, which integrates critical pipelines for data processing, model training, and application deployment. This platform facilitates a unified management system for training and inference tools, accommodating various commercial large language models for private deployment. Furthermore, the platform has been built on a fully China-made computing infrastructure. At peak usage, the platform harnesses computing power equivalent to 1950P.