FastH3-Omni is an accelerated MiniMax H3 Omni Ref variant developed through FastVideo, Nuva Lab, and Reactor. The October 13, 2026 announcement adds image references, audio references, keyframes, and continuation to the FastH3 story. Reactor provides hosted session access. FastVideo subsequently published reference-conditioned weights and training resources on October 20, including separate ComfyUI and local-runtime packages.
This article explains the Reactor announcement, FastVideo confirmation, and Reactor model documentation, reviewed October 14, with release status and pricing rechecked October 21. We have not reproduced speed or quality claims. H3Vid's current online generator uses separately identified H3 models and does not offer FastH3-Omni.
What changes from V2 and Trim?
Nuva Lab describes Omni as its first full accelerated version supporting omni references, in contrast with the earlier text-to-video FastH3 previews. Existing FastH3 V2 and Trim packages remain a different installation and hardware decision. Do not assume a released text-only checkpoint gains Omni input modes through a prompt alone.
H3 Max and H3 Max Turbo are fal's post-trained variants. They are not alternate names for this Reactor deployment. A speed-focused search can surface all three families, so inspect the actual model identity, supported inputs, and provider before comparing timings or copying settings.
The October 20 release now provides downloadable Omni weights. Our ComfyUI and local setup guide explains the required diffusion, encoder and VAE components and the distinct RTX 4090 and 5090 recipes. The FastVideo local guide explicitly says its MLX OmniRef path is not implemented. Consumer-GPU support does not establish Mac or AMD compatibility.
References and keyframes have different jobs
The schema distinguishes appearance references from pinned endpoint frames. Image references guide characters, scenes, and appearance. first_frame_image and last_frame_image instead anchor the opening or closing composition.
| Input | Important rule |
|---|---|
| Image references | Up to three for real-time generation; documentation allows nine when real-time speed is unnecessary |
| Keyframe images | Require at least one ordinary image reference in the same request |
| Audio references | Up to three, each 2–15 seconds; require an image reference or a retained continuation clip |
| Continuation audio | At most two uploaded audio references with continuation |
Keyframes do not consume numbered <Picture N> positions as ordinary references. Label the character and environment with the reference order actually sent, then describe the camera and action. Avoid treating a closing frame as merely another style reference.
An illustrative scene brief might describe a ceramic artist turning a bowl, a slow camera approach, pottery-wheel noise, and rain on a window. Supply a character image when identity matters and an audio example when a particular voice or soundtrack matters. This is a prompt-planning example; it does not establish that a requested identity or sound will be perfectly retained.
Clip lengths and output formats
The overview summarizes approximately 5–15-second clips at 24 fps. The exact accepted durations are 5.875, 8, 10.125, 12.25, and 14.375 seconds. Other values are refused rather than rounded. Do not send an ordinary five-second value based only on the overview's shorthand.
The documented canvases are 16:9 at 1344×768, 1:1 at 768×768, 9:16 at 768×1344, and 4:3 at 1024×768. Images must use supported formats and remain within documented byte, pixel, and aspect-ratio limits. Sandbox conversions do not mean a direct API caller can upload any browser-readable format unchanged.
How the Reactor API works
Reactor connects a session to reactor/fast-h3-omni. A server exchanges its API key for a session token, and a client receives separate video and audio tracks. Once ready, it submits an enqueue command. clip_queued means accepted; clip_generated means ready; a subsequent play command starts delivery unless autoplay is enabled.
That sequence differs from an HTTP job returning a completed MP4. A clip can be generated and queued without playing. Save its identifier if it will become continuation context. Handle generation failure separately from connection failure, and design audio playback around the browser's user-interaction requirements.
Use the official scaffold and tutorial for authentication, uploads, and typed SDK methods. We have reviewed the contract but have not installed or exercised the client. Copying a model slug into an existing fal queue adapter would not supply Reactor's connection, tracks, queue, and recovery semantics.
Continuing a scene
The default latent continuation carries motion, camera, and audio from a generated clip retained in the same session and canvas. It accepts no keyframes. frame continuation uses that clip's final frame as the next first frame and its soundtrack as audio guidance.
Do not pass a separate first-frame image with continuation. A last-frame image requires frame mode and an ordinary reference image. Unknown or stale clip identifiers are refused; a reset clears retained continuation state. Persistent recovery therefore needs more than saving a UUID in a browser draft.
Pricing is session time, not finished-video time
The public pricing endpoint, checked October 21, lists FastH3-Omni at $0.0125 per second, equivalent to $0.75 per minute. The billing documentation defines the meter as session time: it starts at ready and continues while the session reserves compute, including idle time and recoverable disconnections.
An eight-second finished clip is therefore not automatically a $0.10 job. Generation, playback, and any idle connected time can contribute to the session bill. Terminate the session when finished; merely leaving its page or stopping playback is not a reliable budgeting policy.
For local users, choose the released package matching both runtime and hardware, then verify its complete component set. For application builders, verify authentication, termination, recordings, and reconnection before promising a per-clip price. The preview establishes a concrete multimodal access route; performance and local compatibility still need verification on your own device.