StepFun releases Step-Audio 2

Background

StepFun (阶跃星辰) is a Chinese AI company. Step-Audio 2 is one of its speech models and follows the earlier Step-Audio model. A speech model of this kind has to understand what is said and how it is said, and it has to be able to respond in speech when it takes part in a conversation.

What happened

On 23 July 2025 StepFun released the technical report for Step-Audio 2. The report came with two benchmarks: StepEval-Audio-Toolcall, which tests how well a speech model uses tools, and StepEval-Audio-Paralinguistic, which tests how well it handles paralinguistic information, such as tone and manner of speaking.^[1] The report is listed on arXiv under the identifier 2507.16632.^[2]

The next day, on 24 July 2025, demonstration videos were published.^[1] These showed the model working in practice and gave readers a chance to hear it before reading the numbers.

Key details

The open release came in stages. The mini and mini Base versions of Step-Audio 2 were open-sourced under the Apache 2.0 license on 29 August 2025.^[1] A reasoning variant, Step-Audio 2 mini Think, was released on 15 September 2025.^[1] The Apache 2.0 license allows use, modification and redistribution, including commercial use, as long as the license notices are kept.

The report gave the following results. On MMAU, a benchmark of audio understanding, Step-Audio 2 averaged 78.0. On the English basic level of URO-Bench, a test of speech-to-speech conversation, it scored 83.90.^[1] These are the project's own reported figures, and they cover only the benchmarks named in the report.

Impact

Within about two months, from late July to mid-September 2025, the project went from a technical report to open mini models and a reasoning variant. For readers, the release is a concrete example of how a speech model can be described, measured and shared. The two new benchmarks give other researchers a way to test tool use and paralinguistic handling in speech models, although any benchmark measures only part of a model's abilities.^[1] Readers who want to judge the claims for themselves can start from the technical report and the open repository.