StepFun releases Step-2, Step-1.5V and Step-1X at WAIC

Background The World Artificial Intelligence Conference (WAIC, 世界人工智能大会) opened on 5 July 2024. StepFun (阶跃星辰) used the event to release three models from its Step series ^[1]. The release followed the March 2024 preview of Step-2.

What happened At WAIC StepFun released three models. The first was Step-2, the official version of its trillion-parameter Mixture-of-Experts language model. StepFun said Step-2 approached GPT-4 across maths, logic, coding, knowledge, writing and dialogue ^[1].

The second was Step-1.5V, a multimodal model of about one hundred billion parameters that adds video understanding and uses Step-2 to improve its reasoning ^[1].

The third was Step-1X, an image generator. It is built on a self-developed Diffusion Transformer (DiT) architecture and comes in 600M, 2B and 8B parameter sizes. It has been optimised for Chinese elements and culture ^[1]. A diffusion transformer generates an image by gradually removing noise, with a transformer network guiding the process.

Key details Several partners took part in the launch. Cailian Press (财联社) and Guotai Junan (国泰君安) jointly launched 君弘灵犀, a model for securities business. Shanghai Film (上海电影) presented an AI experience based on Havoc in Heaven ^[1].

StepFun also launched a programme called Fanxing, which supports multimodal start-ups and independent developers ^[1].

Significance The event turned the March preview into an official release and added an image generator to the Step series. The performance figures are the company's own claims.

Terms used A foundation model is a large general-purpose model on which other applications can be built. Diffusion models generate images by starting from random noise and removing it in steps, guided by the text prompt. A Diffusion Transformer (DiT) replaces the convolutional networks used in many earlier diffusion models with transformer blocks, the same building block used in most language models.

Why the releases were announced together Step-1.5V's use of Step-2 to strengthen its reasoning means the models were connected in practice, not only by a shared family name. The release therefore presented StepFun as covering several kinds of output at once: text, images and video understanding.