StepFun

StepFun (阶跃星辰) is a Chinese AI company founded in April 2023, according to its own website, which says it develops its own foundation models and has built Step-1, a hundred-billion-parameter language model, and the multimodal Step-1V.¹ Its international developer platform sells API access to its Step language models, led by Step 5 Preview, and to the StepAudio 3 family of models for speech recognition, real-time voice interaction, speech synthesis, voice and music generation.²

Its open-weight releases on Hugging Face began with the OCR model GOT-OCR2_0 in September 2024, which still drew 468,748 downloads in the 30 days to Oct 13, 2026 Measured, followed by Step-Audio speech models in February 2025.³ Step 3.5 Flash, published on Feb 1, 2026, is a sparse mixture-of-experts (MoE) model, one that routes each token through only part of the network, activating 11 billion of its 196 billion parameters per token, with a 256K-token context, under Apache 2.0 Vendor-stated.⁴ It is StepFun’s most-downloaded general language model, with 264,837 downloads in the 30 days to Oct 13, 2026 Measured, as listed in Chinese Open-Weight Models: Downloads, Derivatives and Token Share.³ Step 3.7 Flash, from May 2026, pairs a 196-billion-parameter language backbone with a 1.8-billion-parameter vision encoder and activates about 11 billion parameters per token Vendor-stated.⁵

Its newest flagship has no downloadable weights so far. OpenRouter lists Step 5 Preview as released on Oct 8, 2026, a sparse MoE model with 27 billion active and 600 billion total parameters and a 1M-token context, served by one provider Vendor-stated, and no Step 5 weights appear among StepFun’s Hugging Face releases.⁶³ In its first week on OpenRouter, an API service that routes requests to hundreds of models, it processed 12.09 trillion tokens, third among all models in the seven days to Oct 12, 2026 Measured, as ranked in Most-Used AI Models on OpenRouter by Weekly Tokens.⁷