Step-1

What Step-1 is

Step-1 is the first language model released by StepFun (阶跃星辰), a Chinese AI company. It has about one hundred billion parameters. Parameters are the internal numbers of a neural network that training adjusts; the count is a rough measure of a model's capacity, and a model of this size is what the field usually calls a large language model — a system trained on large amounts of text and used to generate text, answer questions and write code.^[1]

Background

StepFun was founded in April 2023, and it presented its Step series of general-purpose large models at the 2024 Global Developers Pioneer Conference.^[3] Chinese Wikipedia dates the completion of Step-1 to August 2023, that is, several months before the public launch.^[2] The public unveiling took place on 23 March 2024 at the conference in Shanghai, where founder and CEO Jiang Daxin made the announcement during the opening ceremony.^[1][3]

Launch lineup

Step-1 did not appear on its own. It was launched together with Step-1V, a multimodal model, and a preview of Step-2.^[1] "Multimodal" generally means a model takes in more than one kind of input, such as images as well as text, so the launch presented a language model, a model covering other input types, and an early look at the next generation at the same time. The message was that Step-1 was one member of a family rather than a single standalone release.

The performance claim

StepFun said Step-1 comprehensively outperformed GPT-3.5 in logical reasoning, Chinese and English knowledge, mathematics and code.^[3] The announcement gave no specific benchmark figures, so the comparison is a claim by the company rather than a measured result that outsiders could check against published scores. GPT-3.5 is the widely known earlier model from OpenAI, and comparing a new Chinese model against it was a common way for developers to place a system in the field — a reference point, not a full evaluation.

Availability and status

Step-1 is a closed model: it was not put out as downloadable weights for anyone to run and modify, and access was instead controlled by the company. Chinese Wikipedia lists Step-1 as deprecated, meaning it is no longer the model the company points users to.^[2] That is a normal fate for a first-generation system once successors arrive, and it makes Step-1 a marker of a particular moment rather than something in day-to-day use.

Where it sits in the Step series

StepFun said Step-1 was trained successfully before it began work on Step-2, a model with a trillion parameters.^[1] Step-1 therefore functioned as the first rung of the ladder: a model completed in 2023, announced in 2024 together with a multimodal sibling and a teaser of the far larger successor, and retired once that successor line took over. For readers following how Chinese large-model companies formed and released their first systems, Step-1 is a compact example — a company founded in April 2023 shipping a hundred-billion-parameter model, claiming parity or better against GPT-3.5 without published numbers, and immediately moving on to a bigger target.