The name to read carefully

The headline version of this story writes itself: MiniMax's H3 now does synchronized audio, available through an API. That version gets the credit wrong. MiniMax developed the underlying H3 model, but FastH3-Omni is a post-trained variant from FastVideo, the Hao AI Lab project, working with Nuva Lab. Reactor is hosting its initial API preview. Performance claims about this variant should therefore be attributed to Reactor and the post-training collaborators, not presented as independent evidence from MiniMax.

What the variant offers, on the published descriptions, is breadth of input. Inputs: text, image and audio conditioning, including first- and last-frame controls and dialogue generation. The collaborators report support for up to three reference images in real-time mode, or up to nine when real-time speed is not required. The collaborators' 13 October announcement also lists up to three audio references and clip continuation. Reactor's documentation is concrete about the output shape: Reactor's preview API accepts five clip lengths — 5.875, 8, 10.125, 12.25 and 14.375 seconds — at 24 fps, returning separate main_video and main_audio tracks.

A preview with an expiry date

Access is the second thing to read carefully. From its October 13 announcement until the planned October 20 release, FastH3-Omni is available exclusively through Reactor's API. The model weights and training code are not yet publicly available. The preview gives developers an opportunity to evaluate the hosted version, but it also creates a temporary dependency on Reactor's access terms and infrastructure. October 20 is a promised release date, not confirmation that the files have shipped.

The two missing numbers

Reactor describes FastH3-Omni as the fastest and highest-quality open version of real-time H3, but the announcement does not provide a reproducible Omni-specific comparison supporting that superlative. No Omni-specific API price or independent quality benchmark was identified in the available preview materials. Synchronized audio, keyframe conditioning and reference inputs are concrete features developers can test today; the scheduled October 20 weights-and-code release would allow broader evaluation of quality and performance under disclosed hardware and workload conditions. Public speed figures for the earlier FastH3 line exist, but they describe different checkpoints and do not validate the Omni superlative.

One more term belongs in the same check. Open weights are not necessarily unrestricted open source: the FastVideo framework is Apache-2.0, while the H3 weights sit under MiniMax's Community License, and FastVideo's own checkpoint notes state that derived checkpoints inherit that licence. The terms that ship with the October 20 release should be checked before deployment.

What to do before October 20

Treat the Reactor preview as an evaluation opportunity, not a production dependency. Confirm billing and usage limits before running substantial workloads, test representative prompts and retain outputs for comparison. Avoid committing to an architecture that assumes the October 20 open release will arrive on schedule or that future hosting costs will be acceptable.

Sources

  1. [1] Reactor — FastH3-Omni API overview (documentation: accepted clip lengths, 24 fps, separate main_video and main_audio tracks, SDK workflow)Read source
  2. [2] Reactor — Models listing (FastH3-Omni: text, audio and image input; faster-than-realtime generation)Read source
  3. [3] RuntimeWire, "MiniMax brings synchronized audio-video generation to Reactor's API", 13 October 2026, updated 14 October 2026 — announcement report; notes no Omni-specific price or independent quality results in the documentationRead source
  4. [4] Nuva Lab — Dedicated MiniMax H3 & FastH3 video generation (licensing note: H3 and FastH3 use follows the MiniMax H3 license)Read source
  5. [5] FastVideo (Hao AI Lab) — FastH3 checkpoint documentation, Hugging Face (derived checkpoints inherit the MiniMax H3 Community License)Read source