Step-2

What Step-2 is

Step-2 is a language model with a trillion parameters, built on a Mixture-of-Experts (MoE) architecture ^[1]. The distinction matters. In a dense model, every parameter is used to process every token, so cost grows with the total size of the network. In an MoE model the network is divided into many sub-networks, and a routing mechanism activates only a few of them for each token. The total number of parameters can therefore be very large while the computation per token stays far below what a dense model of the same size would need, which is what makes a model at this scale practical to train and serve.

Where it came from

StepFun announced a preview of Step-2 on 23 March 2024, with API trial access for some partners ^[1]. The model was a deliberate step beyond the company's earlier Step-1. Jiang Daxin said Step-2 followed Step-1 because parameter count sets the ceiling on model intelligence, and that scaling to a trillion parameters demands advances in compute, systems, data and algorithms at the same time ^[1]. The distance from preview to full model was short: StepFun said Step-2 reached trillion-parameter scale about 100 days after the March debut ^[2].

The official release

The official version was released at the World Artificial Intelligence Conference (WAIC) on 5 July 2024 ^[2], alongside Step-1.5V and Step-1X ^[4]. Step-1.5V, released at the same time, used Step-2 to strengthen its reasoning ^[2], which shows how the large model was meant to sit underneath other models in the line-up rather than stand alone. Enterprises and developers could apply for access through the StepFun open platform ^[2].

Stated capabilities

StepFun said its systems capabilities greatly improved training efficiency ^[2] — a claim about the engineering around the model as much as about the model itself, since training at this scale is limited by how well hardware, data pipelines and parallelism are organised. The company said Step-2 approached GPT-4 across maths, logic, programming, knowledge, creative writing and multi-turn dialogue ^[2]. Notably, the WAIC announcement gave no benchmark scores for Step-2; the claim was that it felt close to GPT-4 across those areas ^[2].

Later evaluation

In November 2024, Step-2 ranked first among Chinese foundation models on LiveBench, behind OpenAI o1 and Claude ^[3]. LiveBench is a benchmark built from regularly refreshed questions, so a placing there is a snapshot rather than a permanent rating.

Why it matters

A trillion parameters is a scale marker. Models at this size sit at the top of a vendor's line-up, cost the most to train, and are usually the base that smaller or specialised models draw on — as Step-1.5V did for its reasoning ^[2]. StepFun's stated reason for building it was that parameter count bounds how capable a model can become, and that reaching this size required progress in compute, systems, data and algorithms rather than in any single ingredient ^[1]. The roughly 100 days between the March preview and trillion-parameter scale ^[2] also says something about the pace of the field at the time: a preview and a finished model could be separated by a single quarter.

Making the model reachable through an open platform ^[2] placed it in the commercial pattern common to large language models: the largest model is the showcase, and access to it is sold to enterprises and developers who build on top of it.