An end-to-end, chat-driven, never-ending AI video broadcast. Viewers type
!prompt <idea> in Twitch or YouTube chat; an LLM expands each idea into a
styled sequence of scenes; the fast-h3 model generates them as 768p video
clips with synchronized audio; and the stream goes out over RTMP as one
uninterrupted broadcast.
Two halves, one contract:
| Folder | What it is | Runs on |
|---|---|---|
fast-h3/ |
The model: a queue of prompt-driven clip generations with explicit/auto playback. A reactor CLI workspace. |
Reactor Runtime, 8x B200 |
streaming-client/ |
The client: chat → prompt upsampling → scene groups → the model's queue → paced RTMP output. | reactor-sdk (Python), any box with ffmpeg |
They meet on the wire: fast-h3/fasth3_types.py is the client-facing
contract (commands, messages, tracks), and the streaming client speaks
exactly that — through reactor-sdk,
the Reactor Python SDK, which carries the
session, the commands and messages, and the WebRTC media tracks.
The video generator is
FastH3 Preview v1:
MiniMax-H3 (35B) distilled by the
FastVideo project down to four
transformer forwards with 90% sparse video attention, generating video and
audio jointly from text. The model, the distillation, and the inference
engine this repo serves are FastVideo's work — fast-h3/ wraps them in a
queue-and-playout contract for the Reactor platform.
Serve the model (locally with reactor run from fast-h3/, or deploy it),
then:
cd streaming-client
pip install -r requirements.txt # ffmpeg must be on PATH for RTMP
cp .env.example .env # keys, style, sink, chat channel
python main.py --local --sink noop # dry run against a local runtime
python main.py # everything from .envfast-h3/README.md— the model: the queue contract, weights layout, GPU/CUDA prerequisites, performance profile, and deployment learnings.streaming-client/README.md— the client: architecture, sinks and chat sources, moderation, idle filler, presets, and the RTMP/ffmpeg learnings.fast-h3/client/README.md— a minimalreactor-sdksmoke-test client that walks the raw queue contract once.AGENTS.md— the map for coding agents: system picture, load-bearing invariants, and where each kind of change goes. Read it before changing anything.
The code is Apache License 2.0 — see LICENSE and NOTICE. The model weights are licensed separately under the MiniMax H3 Community License, inherited from the base model MiniMaxAI/MiniMax-H3.