Three Companies Built Video Products on MiniMax H3 in One Day. The License Excludes the US.

On September 30, HeyGen, Utopai Studios, and Creatify Labs each announced a video model that had been post-trained on MiniMax's open-weight H3. Within a single day, three vendors had turned one foundation model into three products. The license all three inherited defines Excluded Territories as the European Union, the United Kingdom, South Korea, and the United States.
HeyGen is a Los Angeles company that describes itself as "a domestic corporation established in the United States." Utopai is headquartered in Mountain View, California, and has billed its model as the highest-ranked video model from a US-based company. On the plain text of the H3 license, running or displaying H3-derived output in the United States sits outside the grant, unless MiniMax licensed it separately.
None of the launch announcements mentioned such a license. Whether either company has one is not on the record.
What the license actually says
The H3 license is dated August 2. It defines the Excluded Territories as the EU, the UK, South Korea, and the US. The grant applies "solely within the Applicable Territory," which is everywhere else.
The clause follows where the model is used, not where the licensee is based. Section V.4 bars using, modifying, or displaying the model or its outputs outside the applicable territory. Derivatives, including distillations, inherit the terms with no added conditions, and outputs may not be used to improve other models. Section II invites anyone in an excluded territory to ask MiniMax for a license, and any user above $20 million in annual revenue needs written authorization regardless.
So the constraint has little to do with the nationality of the company. What matters is where the video, images, or avatars produced by the model end up being used. A US-headquartered vendor selling to US enterprise customers would be using the model in an excluded territory.
The three products
HeyGen's catalog page for HeyGen Video 1.0 states that it is "Built on MiniMax H3 and post-trained by HeyGen." Pricing is $0.01 per second through October, half the $0.02 standard rate, available through HeyGen's API, OpenRouter, Runware, and ComfyUI. The quality figures come from HeyGen's own evaluation on Artificial Analysis Arena prompts. HeyGen is on neither AA board, so those are internally normalized numbers, not rankings.
Creatify's Boreal-H3 is post-trained specifically for advertising, with vendor figures that include 85.3% reference fidelity and brief success rising from 28% to 50%. No stated baselines accompany them.
Utopai X is the one with an external ranking. Artificial Analysis's debut post describes it as post-trained on MiniMax H3 and running inside Utopai's PAI platform with no public API. Utopai's own launch post calls the model proprietary and says only that it builds on an unnamed existing foundation model. The H3 attribution is Artificial Analysis's, not Utopai's.

Why MiniMax keeps applauding
MiniMax's behavior does not read like a licensor with a problem. On fal's launch blog in August, the H3 team called fal "a natural partner for H3 Max." Its X account quote-posted HeyGen's launch with "proud to provide the foundation," and did the same for Creatify.
That leaves two readings. Either MiniMax is licensing US use privately and not disclosing it, or the territory clause binds less tightly in practice than it reads on paper. Both are plausible. Neither is confirmed.
The H3 license is not unusual in restricting where a model can be deployed. What is unusual is the gap between the restriction and the way the model is being marketed. A product sold to US enterprises by a US company, built on weights whose license excludes the US, creates a question that a statement from MiniMax would settle in a paragraph. No such statement exists.
There is a reasonable defense on the other side. Territorial restrictions in model licenses are common, and they are often narrower in enforcement than in text, because the vendor's real concern is usually a specific competitive or legal exposure rather than a blanket ban. MiniMax may have granted individual licenses to these companies and simply not publicized it, which would make the whole question moot.
Why post-training on someone else's base pays off
The strategic logic behind all three products is the same, and it is worth naming. MiniMax H3 is a strong open-weight base with unified multimodal context, native stereo sound, up to 2K resolution, and clip lengths around 15 seconds. Building a video model from scratch to that level costs far more than post-training on top of it.
What each vendor adds is domain data and workflow integration. HeyGen uses its own enterprise data to tune output that fits corporate training, onboarding, and advertising briefs. Creatify tunes for product fidelity and readable labels, because ad creative has to keep a product looking like the product. Utopai tunes inside a production platform aimed at scripted, character-consistent work. The base model supplies the capability; the vendor supplies the fit.
The pricing structure follows from that. HeyGen's promotional rate of $0.01 per second, against a $0.02 standard rate, is a customer acquisition window. Compared with other options on HeyGen's own chart, the number sits well below Kling 3.0 Pro at $0.168, Seedance 2.0 at $0.303, and Veo 3.1 at $0.40 per second. Those comparisons are the vendor's, and quality tiers differ, but the direction of the pricing pressure is clear: open-weight bases are pushing the cost of generated video down across the market.
What this means for buyers
For enterprises evaluating these models, the practical question is less about who is right in a licensing dispute and more about what happens to a production pipeline if the question gets answered against the vendor.
The relevant risks are continuity and auditability. If a model's license is later clarified to restrict the territory where its outputs can be used, a vendor may have to renegotiate, re-train, or repackage the product. Customers who built workflows around a specific model endpoint would absorb the disruption. Buyers who need provenance for regulated use cases would need to trace the model lineage with more care than a marketing page usually provides.
There is also a structural lesson in the pattern. Open weights have made foundation models into a base layer that other companies build on. That layer often ships with license terms that were written for a different era of software distribution, where end use was easy to locate. Video models break that assumption. A clip generated on a server in one country can be edited, resold, and displayed anywhere.
Three products built on one foundation model inside a day is a good demonstration of how fast the open ecosystem moves. It is also a reminder that the legal layer has not moved at the same speed. Until MiniMax says which way its territory clause cuts in practice, every buyer downstream is making an assumption.
Related articles
The Eighth Circuit Paused Minnesota's AI Nudification Ban. The Real Fight Is Over Who Is Liable.
The ruling is not a decision that the statute is unconstitutional. It pauses enforcement pending appeal.
California Signed 13 AI Bills in One Day. SB 1000 Deleted the Million-User Threshold.
For a category of startups that assumed they were too small to be regulated, that assumption just expired.
Unitree Open-Sources Its 6B Humanoid Foundation Model: One Model Runs Both Hands and Feet, and Embodied AI Starts Keeping Score by the Job
For robots, the real test comes in those thousands of hours in factories and stores; the launch event is just the opening act.
Claude Sonnet 5.5 Is 30 Percent Cheaper. That Is a Procurement Story, Not a Model Story.
Vendor discount claims are usually measured against a representative workload, and yours is not representative.