The one-line version: on July 30–31, 2026, Shanghai's MiniMax shipped H3, the first credibly open-weight video model that generates 2K clips with synchronized audio in a single pass — and that, not the per-second price, is what changes the decision for a solo founder. Google's Veo 3.1, Kuaishou's Kling 3.0, and ByteDance's Seedance 2.0 are all closed. If you want controllable video-with-sound and the option to eventually run the model yourself, there is now exactly one box to tick.
Don't decide on the per-second sticker#
The first instinct is to line up dollars-per-second and pick the cheapest. Resist it. The numbers you'll find — H3 at roughly $0.13/sec for 2K and $0.09/sec for 768p (MiniMax pay-as-you-go), against Veo 3.1 and Kling 3.0 quoted anywhere from ~$0.03 to ~$0.10/sec in various roundups — are not the same product. Those cheaper figures are usually lower resolution, a "lite" tier, or video without native audio. Comparing them to H3's 2K-with-sound rate is comparing a sketch to a finished frame.
There is also no independent, cross-model quality benchmark of H3 yet. MiniMax claims its video editing ranks first globally and that H3 undercuts mainstream models by two-thirds at 2K (SCMP), but those are vendor claims plus early-tester impressions. So price and quality — the two things you'd normally decide on — are both too fuzzy right now to carry the decision.
When the sticker prices don't compare and the benchmarks don't exist yet, decide on the axes that are durable: can you get the weights, and does it make sound in one pass?
The two axes that actually compare#
Open weights. MiniMax says it will publish H3's weights "in the coming days" under the MiniMax Community License (Reuters). None of Veo 3.1, Kling 3.0, or Seedance 2.0 offer that. For a founder, weights are optionality: on-prem generation for content you can't send to a third party, and a ceiling on per-second cost once your volume is high enough that API pricing dominates. You don't have to self-host on day one — but only H3 lets you ever choose to.
Native audio in one pass. H3 ingests text, image, video, and audio, then generates the picture and synchronized stereo sound together — dialogue, sound effects, and ambient audio arrive with the frames rather than being dubbed on afterward (MarkTechPost). Some Veo tiers do this too; Kling and Seedance are weaker here. For short-form marketing video, one-pass audio deletes an entire pipeline step — the script → render → separate TTS → manual sync dance you're probably doing today.
The license line every founder needs to read#
The Community License reportedly permits free non-commercial use and free commercial use for organizations under US$20 million in annual revenue, with attribution (SCMP). If you're a solo founder or small team, that almost certainly means you can build a real product on H3's weights for free. But "open weights" is not MIT — this is the same revenue-triggered structure that keeps catching builders off guard in the Kimi K3 license. Read the actual text before you commit, and know where your revenue puts you.
So which one do you pick?#
- You need controllable 2K video with native audio, and want the self-host option later → MiniMax H3. It's the only box that ticks all three.
- You're under $20M revenue and want commercial use with no per-second bill → H3's weights (once they land), with the API to prototype now.
- You need proven, top-end cinematic motion today, with a published track record → a closed incumbent (Veo 3.1 for quality-plus-audio, Kling 3.0 or Seedance 2.0 for motion), and accept the lock-in.
- You just need the cheapest short clips at low volume → prototype on whichever API is cheapest at your resolution and audio setting, and re-measure — the sticker prices won't tell you.
The frontier that mattered this week isn't a benchmark. It's that "open-weight, audio-native video" went from an empty category to a shipping product. For anyone whose next feature is video, that's a new option worth holding — even if you don't exercise it yet. (When you do need to serve generative media at scale, our note on where to run it picks up the thread.)
