news

recently, china telecom artificial intelligence research institute (referred to as: teleai) successfully completed the first trillion-parameter large model in china based on the nationally produced wanka cluster training (referred to as: wanka wancan), and officially open sourced the first domestically produced model based on the nationally produced wanka cluster. telechat2-115b, a large model with hundreds of billions of parameters trained by huawanka cluster and domestic deep learning framework, is a large model of star semantics.

this is another milestone and important scientific research achievement led by professor li xuelong, cto, chief scientist of china telecom group, and dean of china telecom artificial intelligence research institute. it marks that domestic large-scale model training has truly realized the substitution of nationalization and officially entered the market. a new stage of independent innovation, safety and controllability for domestic production.

telechat2-115b has been trained based on china telecom's self-developed tianyi cloud "integrated intelligent computing service platform" and the artificial intelligence company's "xinghai ai platform". it uses a variety of optimization methods to improve model training efficiency and accuracy while ensuring training accuracy. stability, achieving more than 93% of the computing efficiency of gpu with the same computing power, while the effective training time of the model accounts for more than 98%.

the open source of telechat2-115b marks another new journey for the localization of large models. as the first state-owned enterprise to lay out and open source large models, teleai actively promotes the continuous progress of large model technology through open source, and continues to promote and lead the rapid transition of technological innovation to industrial implementation.

in the opencampass test list in may this year, the logical reasoning capabilities of the telechat series models ranked first in the list of open source large models. as a new generation version, telechat2-115b ranked first with a score of 86.9 points in the latest c-eval evaluation open access model comprehensive list released in september. its general capabilities are nearly 30% higher than those of the telechat series models, especially in terms of tool use, logical reasoning, mathematical calculations, code generation, and long-form writing.

it is understood that teleai’s self-developed large semantic model has won first place in many authoritative competitions. among them, it won first place in the chinese spatial semantic understanding evaluation and ancient chinese historical event type extraction evaluation at the ccl2024 conference. in addition, he won the championship in the nlpcc2024 chinese argumentative paper mining (shared task5) challenge.

(information)