Another week, another video model launch — that’s basically been the rhythm of 2026 so far. So when MiniMax quietly dropped MiniMax H3 on July 31, I’ll admit my first reaction was mild fatigue. Another resolution bump, another “revolutionary” claim, another benchmark chart nobody outside the lab fully understands.
Except this one’s actually worth stopping for. Not because of the resolution number — though native 2K is genuinely rare — but because of what MiniMax H3 does with footage that already exists. It edits it. Not “regenerate the whole clip and hope it looks similar” editing. Actual, targeted, leave-everything-else-alone editing. That’s the part that made me go back and read the technical notes twice.
Okay, But What Is It Actually
MiniMax H3 — also called Hailuo 3.0, since it’s the third generation of MiniMax’s Hailuo video line — is what the company calls an “omni-modal” model. Fancy term, but the idea is simple enough: it reads text, images, video, and audio together in one pass instead of treating them as separate steps bolted together afterward.
The spec sheet, if you’re the type who wants the numbers up front:
- Native 2K (2560×1440) — not upscaled
- 24fps
- 4 to 15 seconds per clip, stretchable to about 30 with MiniMax’s Extend tool
- Native stereo audio generated alongside the video, not layered on after
- Six aspect ratio options, from 21:9 down to 9:16
None of these numbers alone would get me writing an article. Plenty of models throw around “2K” or “native audio” as marketing lines these days. What’s different here is that they all come out of the same model at the same time — which, if you’ve ever had to stitch together a video tool, an upscaler, and a separate audio generator for one 10-second clip, you’ll understand why that’s not a small thing.
The Omni-Reference Thing Actually Works Better Than I Expected
MiniMax calls it Omni-Reference. You can feed MiniMax H3 up to nine images, three video clips, and three audio clips — all in the same request, all treated as one combined context. A character photo here, a product shot there, a clip showing how you want the camera to move, an audio sample for tone.
In practice, this is the difference between “describe everything in a paragraph and hope the model gets it” and “just show it what you mean.” I’d still call it a mixed bag — reference-heavy generations take real trial and error to dial in, and MiniMax’s own docs are honest that results vary by how clean your source material is. But when it works, it’s noticeably better at holding a character or product consistent than prompt-only generation tends to be.
Here’s the Feature That Actually Matters: Editing
If you only remember one thing from this article, make it this: MiniMax H3 can take footage you already have and change a specific piece of it — swap an object, change the background, adjust lighting, alter how someone’s moving, even replace dialogue with the mouth re-forming around new words — while leaving the rest of the shot exactly as it was.
That’s a genuinely different problem than text-to-video, and it’s reportedly good enough that MiniMax H3 currently sits at the top of the Artificial Analysis video-editing leaderboard, ahead of the other big names in the space. I’d take any single leaderboard with a grain of salt — these things shift fast — but it lines up with what the editing demos actually show.
For anyone doing revision-heavy work — ad variants, product listing tweaks, “can we just change the background color” client requests — this is the feature that saves actual hours, not the resolution number.
Pricing Is the Other Sleeper Story Here
MiniMax priced this to undercut, not match. Around $0.13–$0.14 per second at 2K — which lands below what some competitors charge for plain 720p. If you’re running generations at any real volume, that gap adds up fast across a month, not just a single project.
There’s also an open-weight release reportedly coming “in the coming days” per MiniMax’s own announcement. Worth watching, though I’d hold off calling it a done deal until the weights are actually out — plenty of “coming soon” open releases in this space have slipped.
Who Should Actually Care
Based on what’s shipped so far, MiniMax H3 seems best suited for:
- Ad and brand content where you need fast iteration, not one perfect hero shot
- E-commerce visuals where label and texture legibility matter (2K genuinely helps here)
- UI/UX previews and app walkthroughs — it’s noticeably better at rendering readable on-screen text than most video models
- Game cinematics and pre-vis work
- Anyone doing constant small edits to existing footage rather than generating fresh clips every time
Should You Bother Testing It
Probably, yes — especially if editing existing footage is a real pain point for your workflow, since that’s the one place MiniMax H3 seems to be doing something meaningfully different rather than just incrementally better.
If you want to try it without wiring up a direct API integration, it’s available through the MiniMax H3 AI Video Generator on VidLux, which runs it alongside other video models like Veo, Wan, and Kling — useful if you want to run the same prompt across a few models before deciding what to commit to.
My honest take: the 2K number will get the headlines, but the editing capability is the part that’ll actually change how people work. Resolution races are fun to write about; tools that save you a re-shoot are the ones that stick.
Sources: MiniMax’s official launch blog and Artificial Analysis benchmark results.
