Introduction

Real-time H3 just got ears and a casting department. FastH3-Omni, a post-trained, accelerated take on MiniMax H3's omni-reference model, went live in preview on October 13 from the FastVideo team at Hao AI Lab, Nuva Lab, and Reactor. It generates short video clips with synchronized audio from a text prompt, plus optional reference images, keyframes, and voice or soundtrack clips.

Right now you can run it on the Reactor API (reactor/fast-h3-omni), which holds an exclusive preview until October 20, 2026. That same day, Hao AI Lab says the final model will be fully open-sourced with training code, with support for DGX Spark and RTX GPUs.

What Shipped

Nuva Lab calls this the first full FastH3 release with omni-reference support. Earlier FastH3 v1 and v2 previews only covered text-to-video. Here's what Reactor's docs list for the preview:

  • Output: a main_video track plus a main_audio track, so picture and sound arrive together
  • Clip length: about 5–15 seconds; seconds must be 5.875, 8, 10.125, 12.25, or 14.375
  • Frame rate: 24 fps
  • Canvases: 16:9 (1344×768), 1:1 (768×768), 9:16 (768×1344), 4:3 (1024×768)
  • Image references: up to 3 for real-time generation, up to 9 when real-time speed isn't needed
  • Keyframes: first_frame_image and last_frame_image pin the opening or closing shot (you still need at least one ordinary reference image in the request)
  • Audio references: up to 3 clips, each 2–15 seconds, to guide a voice, ambient sound, or soundtrack
  • Continuation: extend a previous clip in latent mode (carries motion, camera, and sound) or frame mode (reuses the last frame); the session keeps your 8 most recent clips as continuation sources

Reactor's announcement also highlights audio input with "strong dialogue" and first/last-frame conditioning as the headline upgrades.

Open vs. API

Today it's API only, through Reactor. The weights aren't public yet. The open-source drop is promised for October 20, and until then this counts as a preview. Reactor's billing page still marks the FastH3-Omni rate as to be announced, so check your dashboard before you queue a long batch.

Try It Today

Reactor scaffolds a ready-to-run app for you:

npx create-reactor-app my-omni-app --model=fast-h3-omni

Keep your rk_ API key on the server. Reactor exchanges it for a session token, and the browser only ever sees that token. The typed SDK package is @reactor-models/fast-h3-omni.

Paste-ready prompt

Reactor's prompt guide recommends one clear beat per clip, with labeled lines for camera, dialogue, sound, and music. Here's a starter in that format:

A lighthouse keeper climbs the spiral stairs at dusk, lantern swinging, and pauses at the top window to look out at the storm.
Camera: Slow upward tracking shot that ends on her silhouette against the glass.
Dialogue: She whispers, "Right on time."
Soundscape: Wind howls against the tower and the iron steps ring under her boots.
Music: A low cello note swells as she reaches the window.

Reference tips

  • Call your uploads <Picture 1>, <Picture 2>, and so on, in upload order, and give each a job: person, setting, object, or style.
  • Tie a voice to a character in plain words, like "<Audio 1> is the voice reference for <Subject 1>". If an audio clip has no stated role, the model may ignore it.
  • Write the exact words you want spoken or sung in the prompt. A voice reference shapes how it sounds, not what it says.

Why It Matters

FastH3 already made H3-class video feel interactive. Omni adds what real productions need: consistent characters from reference stills, a voice you choose, and clean first and last frames for cutting shots together. If the October 20 release lands as promised with RTX support, that whole kit moves onto local machines.

Original Source

Docs: FastH3-Omni overview on Reactor and the FastH3-Omni prompt guide.

Conclusion

FastH3-Omni is a preview with a date on it. You can build with it on Reactor today, and the open weights and training code are due October 20. Mark your calendar, save your reference stills, and start practicing labeled-line prompts now.

Want to play with MiniMax H3 while you wait? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&amp;utm_campaign=gen&amp;utm_content=fasth3-omni-realtime-h3-video-audio-reactor-open-oct-20

—Aurelia ♡