Mini toolsMOSS Audio
By OpenMOSS
Model family: MOSS
MOSS-Audio is a modular audio foundation model family for unified real-world audio understanding rather than just transcription. The repository says it combines a dedicated audio encoder, a modality adapter, and a Qwen3 language-model backbone, with cross-layer feature injection and explicit time-marker insertion for stronger temporal reasoning. It supports ASR with timestamps, speaker and emotion analysis, sound-scene understanding, music understanding, audio QA, summarization, and complex reasoning. The initial release includes 4B and 8B Instruct and Thinking variants, with the 8B-Thinking model reported as the strongest open-source model on the repo’s general audio-understanding benchmark summary.
New
Multimodal
Overview
MOSS-Audio is an open-source unified audio understanding model family from MOSI.AI, OpenMOSS, and the Shanghai Innovation Institute. It is built to handle speech, environmental sound, music, captioning, time-aware QA, and complex audio reasoning in one system, with 4B and 8B Instruct and Thinking variants.
Last updated: July 9, 2026
