StepFun debuts with the Step-1, Step-1V and Step-2 preview models

Background StepFun (阶跃星辰) had been developing its models before this debut. The company's founder and CEO, Jiang Daxin (姜大昕), presented the Step series at the event.

What happened On 23 March 2024 StepFun made its public debut at the 2024 Global Developers Pioneer Conference (全球开发者先锋大会) in Shanghai ^[1]. Founder and CEO Jiang Daxin announced the Step series. It comprised Step-1, a language model with about one hundred billion parameters, and Step-1V, a multimodal model of the same size. It also included a preview of Step-2, a trillion-parameter language model that uses a Mixture-of-Experts (MoE) design ^[1].

Key details Step-1V ranked first on the OpenCompass (司南) multimodal leaderboard. StepFun said its performance was on par with GPT-4V ^[1].

The trillion-parameter Step-2 preview was offered through an API, and some partners were given access to trial it ^[1].

Two consumer apps were made fully available. Yuewen (跃问) is a productivity tool. Maopaoya (冒泡鸭) is an AI open-world platform ^[1].

Several partners were announced at the event. Cailian Press (财联社) would provide AI financial news and investment research. China Literature Online (阅文集团) and CNKI (中国知网) would provide literature and knowledge services ^[1].

Significance The debut set out the company's plan in a single announcement: language models, a multimodal model, a larger model still in preview, and consumer and partner applications.

Terms used at the debut A multimodal model accepts more than one type of input, such as images together with text. A preview is an early version offered to a limited group so that developers and partners can test it before a full release. An API (application programming interface) is a standard way for one piece of software to send requests to a model running elsewhere, so a partner can try a model without running it on its own machines.

Reading the launch The debut put three kinds of product side by side. A language model and a multimodal model are general-purpose building blocks that other software can call. Consumer apps are the form in which ordinary people use models directly. The partners named at the event point to specialised uses: financial news and investment research, and literature and knowledge services.