MOSS-Audio

OpenMOSS's unified audio-understanding model family — speech, environmental sound, music, captioning, and time-aware audio QA/reasoning in one model (8B and 4B, Instruct and Thinking). Built on a Qwen3-8B backbone with a from-scratch MOSS-Audio-Encoder and DeepStack cross-layer injection. The 12.5 Hz MOSS-Audio-Tokenizer alone has 260K+ downloads; the 8B variants ~35K each — one of the highest-traction academic audio releases of 2026. From the OpenMOSS team (now migrating to "OpenMOSS (SII)" / MOSI.AI branding; filed with the rest of Fudan's MOSS line).