MiniMax H3: The New AI Video Model Nobody Outside China Is Talking About Yet

So there’s a new video model out and I honestly wasn’t expecting to be this impressed by it. MiniMax H3 launched out of Shanghai this week, and unlike most “big AI announcement” posts, this one actually has the specs to back up the hype: 2K video, real stereo sound baked in, up to 15 seconds a clip, and — this is the part that got my attention — open weights dropping within days so anyone can download and run the thing themselves.

What H3 Actually Does

Here’s the thing that makes MiniMax H3 different from most video models I’ve written about: it doesn’t treat text, images, video, and audio as separate jobs. You can throw up to 9 reference images, 3 video clips, and 3 audio tracks into a single generation, and the model reads all of it as one connected context — identity, camera movement, composition, sound, editing rhythm, all of it — then carries that through to a coherent result.

In plain terms, that means you can hand it a face from one photo, a camera move from a reference clip, and a voice from an audio file, and it’ll blend all three into a new shot instead of you juggling five different tools. It also does proper editing on footage you already have with MiniMax H3 — swap a product, change signage text, relight a scene from day to night, add or remove an object — and it keeps everything around your edit stable instead of regenerating the whole frame from scratch.

Oh, and it actually renders legible text. Subtitles, credits, brand logos, UI elements for things like game menus or landing pages — MiniMax says it handles all of that cleanly, which has been a genuine weak spot for a lot of video models up to now.

Why “Open Weights” Is a Bigger Deal Than It Sounds

Most of the flagship video models we cover — Veo, Sora when it existed, Seedance — are closed. You use them through an API or an app, full stop. MiniMax is doing something different: they’ve said the H3 model weights will be released within days, meaning anyone with the hardware can download it, self-host it, and modify it however they want.

That’s a real shift, not a marketing line. It puts H3 in the same camp as other Chinese AI labs that have been pushing open weights hard this year, and it means smaller studios, indie developers, and researchers get access to genuinely frontier-level video generation without being locked into someone else’s pricing or rate limits. Whether you’ll actually want to self-host it is another question — most people will still just use the hosted API — but having the option changes the power dynamic a bit.

The Price Is the Real Headline

Okay, this is the number that actually matters if you’re thinking about your own budget. MiniMax says generating 2K video with H3 costs less than a third of what mainstream rivals charge, .and MiniMax H3’s lower 768p tier comes in under half the price, it’s under half the price of typical 720p output elsewhere. Early third-party pricing trackers put the 2K rate around $0.13 per second, which works out to roughly $1.95 for a full 15-second clip — though it’s worth flagging that figure is reported rather than confirmed directly on MiniMax’s own pricing page as of this writing.

Either way, that’s a meaningful undercut. If you’ve been holding off on a video tool because the per-generation cost adds up fast when you’re testing a lot of ideas, this is the kind of pricing that changes the math.

How It Stacks Up Against Seedance and Gemini

I’m not going to pretend H3 wins across the board, because it doesn’t. According to benchmarking from Artificial Analysis, H3 actually leads the field specifically on video editing, but it trails Gemini Omni Flash on straight text-to-video, and sits behind both Seedance 2.0 and Gemini Omni Flash on image-to-video quality. So this isn’t a “best model ever, period” story — it’s a “genuinely strong option that’s exceptional at one specific thing and priced aggressively everywhere else” story.

Runway has already picked H3 up too, adding it alongside Seedance 2.0, Kling 3.0, and Sora 2 Pro in its own multi-model platform, which tells you the industry is taking this seriously rather than writing it off as another minor China-based release.

Should You Bother Trying It?

If editing existing footage is a big part of your workflow — swapping elements, relighting scenes, fixing text without re-rendering everything — MiniMax H3 is genuinely worth testing given it’s apparently leading the pack. If you’re mostly doing pure text-to-video generation, Seedance 2.5 or Gemini Omni Flash are still probably your stronger picks based on current benchmarks.

Either way, keep an eye on this one. A model this capable, priced this aggressively, with open weights landing any day now — that combination tends to shake up a market fast, and I’d rather you hear about it from here before it’s suddenly everywhere.

Want more of what’s actually happening in AI video right now? Check out our coverage of why OpenAI shut down Sora, Meta’s rocky Muse Image launch, and Google making Veo 3.1 free for everyone for the bigger picture on where this space is heading.