At the WAIC 2024, StepFun unveiled three new products in the Step series of universal large models: Step-2 trillion-parameter language large model official version, Step-1.5V multimodal large model, and Step-1X image generation large model. These products aim to improve training efficiency and approach GPT-4 in various areas such as mathematics, logic, programming, knowledge, creativity, and multi-turn dialogue. StepFun also introduced self-developed large model application products for C-end users and discussed their latest developments and plans for large model ecological cooperation.
The Step-2 trillion-parameter language model official version features an innovative MoE hybrid expert model architecture and can be experienced through StepFun’s open platform. Additionally, StepFun announced the launch of an open platform called “Starry Plan” to support startups and independent developer teams in the multimodal field, accelerating the application of AI large model technology in different scenarios.
Founder and CEO of StepFun, Jiang Daxin, emphasized the importance of “trillion parameters” and “multi-modal fusion” in reaching the peak of AGI. He stated that trillion parameters are essential for achieving AGI, and multi-modal large models are the key to AGI. Moving forward, StepFun aims to create even bigger and stronger models to serve both work and life, expanding possibilities for everyone.
Source link
Source link: https://pandaily.com/stepfun-releases-three-large-models-of-the-step-series/