Yichi Zhang

Core Dev @ SGLang Omni

[Yichi Zhang @YichiZ03]

MOSS-Transcribe-Diarize-0.9B is out — congrats to @MosiAI_Official 🎉

Two things stand out:
90 min of audio in a single pass, and multi-speaker transcription.

Both push hard on the serving stack. Day-0 support in SGLang-Omni — run it today on cookbook: https://t.co/6IHfhr2D27

[MOSI @MosiAI_Official]

🤗 MOSS-Transcribe-Diarize-0.9B is now open source on @huggingface.

Built with an end-to-end audio-to-structured-transcript paradigm:

>0.9B open-source ASR model
>Apache license 2.0
>128k long-context transcription
>Up to ~90-min audio input
>Speaker labels + timestamps in one generation
>Multi-speaker diarization for meetings, interruptions, and overlapping voices
>Hotword biasing for names, terms, and domain-specific vocabulary
>~100 token/s on NVIDIA RTX 4090, RTF ~0.017

Thank you @sgl_project @vllm_project @Prince_Canuma @lllucas for day-0 support! 🚀

Github: https://t.co/0hgIjvhGv1
Huggingface: https://t.co/gQ1avC3w2e
API: https://t.co/znzfO4PHuu
Live demo: https://t.co/wAmEbbFna5
Technical Report：https://t.co/3yb0VKp7FS
HF Space: https://t.co/gQ1avC3w2e
AtomGit：https://t.co/uIwqnkMeGU
SGLang-Omni: https://t.co/TFlp4YzI7f
vLLM: https://t.co/j1VMVuQDcp
MLX-audio: https://t.co/ofDO9V6N7J
Discord:https://t.co/H3shzNLHOL

[Yichi Zhang @YichiZ03]

People may think serving and optimizing TTS (Text to Speech) model is similar to traditional LLM. It isn't.

We made Higgs and MOSS-TTS 2–3.4× faster in SGLang-Omni — and most of the wins landed outside the LLM backbone.

Delay patterns, D2H stalls, a vocoder tail-latency killer... In this post, we share everything we learned — every optimization, and every trade-off behind it. We hope it proves a valuable resource for developers building in this space.

Full breakdown:
https://t.co/bTttx61ysc

@sgl_project

[Yichi Zhang @YichiZ03]

https://t.co/o9aPp3UjK1

[Yichi Zhang @YichiZ03]

https://t.co/o9aPp3UjK1

[Yichi Zhang @YichiZ03]

A year ago when I just graduated, I thought AI was overhyped and thought I'd never touch it.

Today I'm shipping AI infra for SGLang-Omni.
I wrote up everything I figured out — why I switched, what AI infra actually does, and the exact methods that got me from zero to landing real PRs on a huge open-source codebase.

If you're a SWE eyeing this space, this is the guide I wish I had and hopefully useful to you:
https://t.co/I37SBV28j4

[Yichi Zhang @YichiZ03]

https://t.co/8tDLheGgsV

[Yichi Zhang @YichiZ03]

https://t.co/8tDLheGgsV