(Yicai) Feb. 18 -- Geely Holding Group's car manufacturing unit and artificial intelligence startup Stepfun will make two of their text-to-video and text-to-speech models open source, following in the footsteps of industry pioneer DeepSeek.

The source code of Step-Video-T2V, a video generation model that can produce high-quality videos at 540p resolution with 204 frames, and that of Step-Audio model, which can create speech with diverse emotions and languages, will be shared with global developers, Hangzhou-based Geely Auto Group and Shanghai-based Stepfun jointly announced today.

The ecosystem partners are reinforcing a trend of reduced secrecy while making large language models accessible for community iterations, inviting improvements. The trend was initiated by Hangzhou-based DeepSeek, whose powerful and affordable DeepSeek-R1 model, launched in January, has disrupted the AI landscape. Even search engine giant Baidu announced last week that its next version of Ernie Bot will be open-source from June 30.

Established in April 2023, Stepfun has collaborated with the owner of Geely Auto, Lynk & Co and Proton marques on application scenario design, model evaluation, and engineering. Stepfun's founder Jiang Daxin formerly served as vice president and chief scientist at Microsoft Software Technology Center Asia, a research and development hub that focuses on machine learning, cloud computing, and Big Data.

Geely has deeply integrated its Xingrui model, released in January, with the Step series, supporting the development of smart driving and intelligent cockpit features, Chief Executive Officer Gan Jiayue said today. For instance, Geely is using AI-generated scenarios for smart training, majorly enhancing autonomous vehicles' ability to adapt to diverse road conditions.

The Chinese auto giant is also leveraging DeepSeek's open-source LLM. On Feb. 7, Geely announced that it has deeply integrated its self-developed Xingrui model with DeepSeek-R1, enabling the smart cockpit of its vehicles to accurately understand even users' vague intentions, significantly improving the user experience.

Editors: Dou Shicong, Emmi Laine