← sources

MOSS AMA snipits

source: reddit (reddit:StableDiffusion)published 2026-10-31 04:40ingested 2026-11-01 02:38

retained text

In Progress:

* [https://mosmos.moss.cn/en.html](https://mosmos.moss.cn/en.html)
*  next-gen full-duplex end-to-end conversational model, [MOSS-Live](https://www.moss.cn/products/moss-live) (similar to GPT-Live)

Soon:

* MOSS-Anime
* Image "Turbo's on the way" (sounds like later-soon IMHO)

Someday later:

* Image edit / Ref2Img
* **smaller Text2Vid** model: "audio-video alignment and motion capability (motion may be a relatively difficult part). We will train with the currently relatively small parameter count, with the hope that it can run on consumer-grade GPUs.", "currently at roughly the mid-stage of training", "we trained the VAE ourselves", "moss-image is \[was\] the first step"
* update the Voice Generator capability in the next-generation TTS-2.0
* add audio input to moss-vl and moss-vl-realtime
* "MOSS’s Audio Understanding team is currently primarily developing our next-generation multi-speaker speech recognition model, MOSS-Transcribe-Diarize"

Info

* trained MOSS-Image on Alibaba's in-house PPU (Parallel Processing Unit) series (from T-Head)
* " I guess open source really brings people together around that culture of mutual support"
* [MOSS-Image Technical Report](https://arxiv.org/abs/2610.23374)

document fetched 2026-11-01 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xk14hd/moss_ama_snipits