Models

MiniMax H3 (Hailuo 3.0): Siray's New Open-Source Video Model, Explained

Siray.AI

31 Jul 2026 • 3 min read

MiniMax H3 (Hailuo 3.0): Siray's New Open-Source Video Model, Explained

Key Takeaways

MiniMax H3 (Hailuo 3.0) is an open-weight video generation model, now live on Siray as of July 31, 2026.
It generates 2K/24fps clips 5-15 seconds long, extendable to roughly 30 seconds with its Extend tool.
Omni-reference locks style, character, motion, and voice from up to 9 images, 3 videos, and 3 audio clips in a single call — with native stereo audio baked in.
MiniMax positions it as a general commercial creative tool for ads, e-commerce, games, and UI motion — not a niche or NSFW-specific model.

Introduction

MiniMax H3, also known as Hailuo 3.0, previewed at WAIC in mid-July and went live on Siray on July 31, 2026. It's an open-weight video model, which means the underlying architecture is publicly documented rather than locked behind a single vendor's black box — and Siray now hosts a ready-to-call endpoint for it, so nobody has to stand up their own GPU cluster to use it.

The model is aimed at teams that need commercial-grade video output fast: ad creatives, e-commerce product shots, game trailers, UI motion mockups. If that's the job, MiniMax H3 is worth a look.

Overview: What MiniMax H3 Actually Does

MiniMax H3 outputs at 2K resolution and 24fps, with a native clip length of 5-15 seconds. Its Extend tool can stretch a generation out to roughly 30 seconds when a scene needs more room to breathe. Aspect ratio is flexible across the range from 21:9 down to 9:16, covering both widescreen and vertical/social formats from the same model.

The feature that sets it apart is omni-reference: a single generation call can take up to 9 reference images, 3 reference videos, and 3 reference audio clips, and use them to lock style, character appearance, motion, and voice simultaneously. Most video models handle one reference type at a time — an image for likeness, maybe a separate step for motion. MiniMax H3 folds all of it into one request.

The model also generates native stereo audio alongside the video, supports instruction-guided edits (describe a change in plain language rather than re-prompting from scratch), and can do video-to-video motion transfer — taking motion from a reference clip and applying it to a new subject.

Why It's Different from a Typical Video Model

Most video generators on the market still treat audio as a separate pipeline step, bolted on after the visual is done. MiniMax H3 generates stereo audio in the same pass as the video, which cuts out a whole post-production round trip. Combined with omni-reference, that means a single API call can produce a clip that already has consistent character appearance, matching motion, and synced sound — no compositing required.

Where It Fits: Two Quick Use Cases

E-commerce teams can use MiniMax H3 to turn a handful of product photos into a launch-ready video ad without hiring a video team for every SKU — a fuller playbook for that workflow is coming in a follow-up article on product launch videos.Game and UI teams have a different problem: short motion loops for trailers, menu transitions, or interaction previews. MiniMax H3's short native duration and instruction-guided edits fit that iteration cycle well — a dedicated walkthrough on that use case is also coming.

Developer Notes

MiniMax H3 is available through Siray's unified API, so teams already calling Siray for other models (LTX 2.3, Seedance 2.0, and others — see One API for Multiple AI Models) don't need a separate integration. For a from-scratch walkthrough of calling a Siray video model, see the AI video generation API quickstart.

Summary

MiniMax H3 is live on Siray today: 2K/24fps, 5-15 second clips extendable to ~30 seconds, omni-reference across images/video/audio, native stereo sound, instruction-guided edits, and video-to-video motion transfer. It's built for general commercial use — ads, e-commerce, games, UI — not locked to one niche.