实时视频模型
MiniMax H3
开源权重 33B 视频模型(Hailuo 3.0)——15 秒带原生立体声片段;这波浪潮的地基。
huggingface.co · 收录≠背书 · 2026-09-03 复核
- 参数
- 33B dense · open base weights (H3-Base)
- 输出
- 4–15s clips; 2K via API only (local ≈768p)
- Audio
- Native stereo — a property of the model
- 开源时间
- Aug 3, 2026
- 许可说明
- Regional exclusions incl. US / EU / UK / KR
什么是 MiniMax H3?
这波实时生成浪潮脚下的开源 33B 视频模型(Hailuo 3.0)。只有 base 权重开源(本地约 768p),2K 与参考模块走 API,且授权有区域排除。MiniMax 官方 API 不是实时的——由 fal 的实时化版本把它变成流媒体基底。
MiniMax H3 能做什么
- 33B dense open-weight model (H3-Base) released Aug 3, 2026
- Generates 4–15 second clips with native 32kHz stereo audio — audio is in the model, not post-processed
- Omni-modal: text, image and video in; video and audio out
- 2K upscale and context-reference modules are API-only — local open-weight generation is around 768p
- Day-zero ecosystem support from Hugging Face, ComfyUI, ModelScope and many chip vendors
MiniMax H3 适合谁?
- Self-hosting open-weight video generation on a local GPU
- Building video pipelines that need native, controllable audio
- Running realtime or near-realtime generation via fal’s H3 Max build
MiniMax H3 的优点与不足
优点
- One of the first flagship open-weight video models with native audio
- Runs locally — ComfyUI shipped day-zero support and cites RTX-3060-class GPUs
- Day-one adaptation across major open-source tooling and silicon vendors
不足
- Only the base is open; the 2K pass and reference modules are API-only
- Licence carries regional exclusions (incl. US, EU, UK, South Korea)
- MiniMax’s own API is not realtime — the streaming use case runs through fal’s tuned build
别人怎么说
摘自公开报道的原文引用——收集于 2026-09-03。
“MiniMax H3 dropped today with open weights, and it’s natively supported in ComfyUI as of this morning. Day zero.”
“Audio is a property of the model, not a post-process. Every audio output is native stereo.”
常见问题
Is MiniMax H3 open source?
The H3 base weights are open (Aug 3, 2026); the 2K regeneration and context-reference modules are API-only.
Can I run MiniMax H3 locally?
Yes for the open base — around 768p resolution; 2K output needs the API modules.
What does native audio mean for H3?
Audio is generated as a property of the model — native stereo — rather than added as a post-processing step.
Is MiniMax H3 realtime?
Not on MiniMax’s official API. fal’s H3 Max build post-trains it to generate faster than playback for live use.
fal H3 Max Live
fal 实时调优的 H3 Max——5 秒 768p 视频不到 3 秒生成,快于播放。fal.ai
Krea Realtime 14B
Open 14B autoregressive model for real-time, long-form streaming video — distilled from Wan 2.1, ~1s to first frame.krea.ai
Decart MirageLSD
“Live-stream diffusion” — restyle a live feed (camera, game, call, screen) frame-by-frame from a text prompt at near-zero latency.decart.ai
Lightricks LTX Video
Open latent-diffusion video model that generates clips faster than real time — the engine several chat-driven live loops run on.github.com