On May 15, according to the "Interim Measures for the Management of Generative AI Services", the Cyberspace Administration of China released a list of 19 new generative AI services that have been filed in Beijing, including 4 star enterprise models: Xiaomi model, Tencent Cloud industry model, DeepseekChat model, Orion Star model, etc., which marks that enterprises can officially provide generative AI services to the public.
It is worth noting that the Orion Star model was released at the beginning of this year, which is the latest achievement of Orion Star after All in AI in 2016, positioning it as "born for enterprise applications". With its 14 billion parameter scale, the model shows excellent performance, not only leading the industry in the base capacity of tens of billions of parameters, but also showing the highest level of SOTA in the same level of models on third-party test sets such as MMLU, C-Eval, CMMLU, GAOKAO, BBH, etc., and can achieve the model effect of hundreds of billions of parameters when combined with enterprise private data and applications.
The "100 model war" is in full swing, but the Orion Star Model has taken a different and pragmatic path. Different from the mainstream industry model of "spending money to train hundreds of billions of models, and then looking for large model application scenarios", Orion Star has always adhered to the thinking logic of Think Different, choosing "born from the application and co-creating with customers", first mining enterprise application scenarios, then defining vertical models, polishing the model based on enterprise private data, and finally creating a cost-effective model.
Orion Star strives to build an "AI technology tree", on the basis of realizing large model capabilities, from the underlying cloud management services, the suite family bucket and turnkey services at the model and suite layer, to the AI digital employee + large model robot at the application and terminal layer, and the AI boss think tank, and finally complete the landing of the customer layer, the technical architecture is closely related and progressive, providing customers with a closed-loop large model application experience.