VibeusSpeech, sound, music, images and video: the MOSS team has opened an... - Vibeus
Speech, sound, music, images and video: the MOSS team has opened an AMA covering its whole model family.
The Shanghai-based team grew out of OpenMOSS research at Fudan University and the Shanghai Innovation Institute. Its broader goal is AI that can follow speech, visual information and changes in a situation over time, spanning generation, understanding and synchronized audio/video.
🎨 Images
MOSS-Image uses a 5B DiT backbone, supports Chinese and English text in images, and accepts JSON layouts for composition control. Its base weights, SFT weights and included inference code are available under Apache 2.0. In the AMA introduction, the team reports 3,702,490 downloads for the ComfyUI version and a place among Hugging Face’s trending image models.
🎙️ Speech and sound generation
• MOSS-TTS handles synthesized speech and cloned voices.
• MOSS-TTSD produces dialogue with multiple speakers, including longer conversations and podcasts.
• MOSS-TTS-Realtime streams generated speech for conversational use.
• MOSS-TTS-Nano is a 100M-parameter option designed for lightweight, CPU-friendly streaming.
• MOSS-VoiceGenerator builds a voice from a written description, without requiring a reference recording.
• MOSS-SoundEffect produces sound effects from text.
🎧 Understanding audio
• MOSS-Transcribe-Diarize combines transcription with speaker labels and timing.
• MOSS-Audio interprets speech, environmental audio and music.
• MOSS-Music analyzes music, including lyrics, instruments and structure.
🎬 Vision and video
• MOSS-VL understands images and video, including long recordings.
• MOSS-VL-Realtime works with a video stream as it unfolds.
• MOVA generates synchronized video and audio.
The source collects the model links, API/playground, Hugging Face and GitHub resources. The live question window was October 30, 8–11 AM PT; the team said it would keep checking questions for another 48 hours. Participants include garybryant, FarAdhesiveness9577, Phospheneser, TimeEnvironmental219, Aggressive-Sign-8167 and Professional-Milk422.
A few development updates from the team’s replies: MOSS-TTS-2.0 is targeting November, with style and emotion control still being refined. Image-editing and reference models are in development, with the control format still experimental. A small-parameter experimental video model is around the middle of training. These are development plans, rather than announcements that those models are already available.
