StepFun launches Step-2, a trillion-parameter MoE model
StepFun said the mixture-of-experts model approximated GPT-4 on maths, logic, coding and dialogue; it was unveiled alongside a multimodal and an image-generation model at WAIC.
- Models & capabilities
- Open weights & ecosystem
- Minor
Chinese startup StepFun unveiled Step-2, a trillion-parameter mixture-of-experts language model, at the World Artificial Intelligence Conference in Shanghai on 4 July 2024, alongside a multimodal model (Step-1.5V) and an image-generation model (Step-1X). It was among the first trillion-parameter MoE models released by a Chinese AI startup rather than a large tech incumbent.
StepFun said Step-2 “approximates GPT-4” across mathematics, logic, programming, general knowledge, creative writing and multi-turn dialogue — a company self-comparison rather than an independently verified result. The model later placed among the top-ranked Chinese models on the LiveBench benchmark, trailing only OpenAI’s and Google’s leading systems globally, according to third-party tracking reported months after launch.
The release fit a broader pattern of well-funded Chinese AI startups — alongside DeepSeek, Moonshot AI and Zhipu AI — pursuing trillion-parameter scale using mixture-of-experts architectures to control inference cost, while competing on domestic benchmarks and increasingly on international ones as well. StepFun went on to raise further funding through 2024 and 2025 on the strength of the Step series.