Category · the models
Realtime video models — the engines of live AI
The "realtime" label is the whole difference. Most AI video models generate offline, taking minutes for a clip you then upload. A realtime video model is post-trained for speed — today that means fal’s H3 Max build of MiniMax H3, which fal reports at roughly 35× the throughput of MiniMax’s official endpoint, producing about five seconds of 768p video in under three.
Listed · 6
The generation work that makes faster-than-playback video possible.
MiniMax H3
Open-weight 33B video model (Hailuo 3.0) — 15s clips with native stereo audio; the substrate under this wave.huggingface.co
fal H3 Max Live
fal’s realtime-tuned H3 Max — five seconds of 768p in under three, faster than playback.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
ByteDance Seed Seaweed APT2
Real-time interactive video research — an 8B model that emits a latent frame per forward pass at 24fps.seaweed-apt.comBenchmarks, with a caveat
On Artificial Analysis, H3 Max led image-to-video-with-audio at Elo ~1,204 as of early Sep 2026; on text-to-video it sits lower. The speed and throughput numbers are fal’s own and have not been independently reproduced — another outlet cites up to 50×. Resolution tops out at ~768p; MiniMax H3’s 2K regeneration pass is API-only.
The open-weight base
MiniMax H3 (Hailuo 3.0) is the open-weight 33B base underneath the wave — 15-second clips with native stereo audio, weights opened Aug 3 2026. Only the base is open (locally around 768p); the 2K and reference modules are API-only, and the licence excludes several regions including the US and EU.
Frequently asked questions
What is a realtime video model?
A video model that generates clips fast enough to sustain a live stream — faster than playback, rather than an offline render job.
Which model powers the 2026 live-AI wave?
fal’s H3 Max Live, a realtime-tuned build of MiniMax H3 — used by fal.live and InfiniteSlop.
Is MiniMax H3 open source?
The H3 base weights are open (Aug 3, 2026); 2K regeneration and reference modules are API-only, and the licence has regional exclusions.
How fast is H3 Max, really?
fal reports ~35× official H3 throughput and ~2.8s for a 5s/768p clip; those figures are vendor-reported and not yet independently reproduced.