Step-Audio 2

Step-Audio 2 mini was released by StepFun on August 29, 2025 as the first fully open-source (Apache 2.0) speech-to-speech multimodal LLM. The 8B-parameter model ingests raw audio tokens and emits audio directly, preserving emotion, accent, timbre, and background context without the fragile ASR-to-LLM-to-TTS pipeline. It supports multilingual conversation, tool calling, web search, and voice-style switching.

It reports state-of-the-art results across speech recognition, audio understanding, and speech conversation benchmarks, including 83.1% on StepEval-Audio-Paralinguistic versus 43.5% for GPT-4o Audio (vendor-reported; treat as vendor claims). Weights, code, benchmarks, and a technical report are all published, with the larger research checkpoint also shipped. A later sibling, Step-Audio-EditX (3B), covers expressive audio editing.