StepFun

StepFun was founded in 2023 by Jiang Daxin, a former Microsoft vice president and chief scientist, and is based in Shanghai.^{1, 2, 4, 5}

The company describes itself as dedicated to advancing AGI, and its work is organized around the Step series of foundation models.^1, 3

By June 2024 it had launched the multimodal Step-1V and was testing Step-2V, and in July 2024 it presented the official version of Step-2, a language model with one trillion parameters.^2, 5, 12

In February 2025, together with Geely, StepFun released the weights of the Step-Video-T2V video model and the Step-Audio speech model.^{4, 13, 14}

Step 3, a multimodal model with 321 billion parameters, followed in July 2025 with weights under the Apache 2.0 license.^{15, 16, 17}

In January 2026 the company announced Yin Qi as chairman, and Jiang Daxin is chief executive.^6, 7

It describes foundation models and AI-enabled terminals such as phones and cars as its long-term strategy.^1, 6

Step 3.5 Flash, released as open source on 2 February 2026, activates 11 billion of its 196 billion parameters per token.^18, 19

Step 3.7 Flash followed on 29 May 2026, also released as open source.²⁰

On 20 September 2026 StepFun launched Step 5 Preview, with 600 billion total and 27 billion active parameters and a context window of up to 1 million tokens.^{9, 21, 22}

Its weights were published on 15 October 2026 under the StepFun Model License 1.0 rather than Apache 2.0.^{3, 8, 9, 21, 22}

The StepAudio family of speech models sits alongside the Step language models.^1, 3