Chinese engine, US badge: How MiniMax H3 quietly powers the top of AI video
Look under the hood of today's video leaderboards and you'll find the same Chinese model almost everywhere: MiniMax H3. US companies like HeyGen, Pika and fal sell post-trained versions under their own names, some of them cheaper than the original. We trace the family tree and work out what's legal and affordable if you're in Germany.

What's going on
A year ago, the leaderboard of the best AI video models was firmly in American hands. In November 2025, US models held the top five spots in the text-to-video arena, led by Google's Veo 3.1. Today the picture looks very different. In the "text-to-video with audio" leaderboard from independent benchmarking firm Artificial Analysis (AA), Alibaba's Wan 3.0 leads with 1,174 Elo points. Elo is a rating system borrowed from chess, fed by blind comparisons in which people pick the better of two videos. Behind it come Utopai X (1,168), MiniMax H3 (1,159) and ByteDance's Seedance 2.5 (1,156). Korea's AI Times calls it a "reversal within a single year." According to AI Times, OpenAI already shut down video generation with Sora in March 2026. Veo 3.1 now sits at just 963 points.
But the real story is one layer deeper. Many of the names that sound like homegrown American models are actually post-trained versions of the same Chinese model: MiniMax H3. We trace the family tree, compare prices and work out what's allowed in Germany.
The family tree: one open model, many brands
MiniMax released H3 on August 3, 2026. The freely downloadable part is the base version with a 33-billion-parameter transformer. It generates video with stereo audio at 768p, 4 to 15 seconds long, at 24 frames per second. According to the Model Card, two components remain closed: a prompt preprocessing stage (H3-Context-IR) and 2K upscaling (H3-Regenerate-2K). That's why DeepLearning.AI's The Batch called the model "only minimally open."
Even so, the derivatives are plentiful. H3's Hugging Face page lists 228 finetunes, 200 adapters, 90 quantizations and 12 merges, 530 derivatives in total. Quantizations are slimmed-down versions that store numbers at lower precision. The main repo reports just under 5.9 million downloads last month. The repackaged version from Comfy, the team behind the ComfyUI interface, even hits 33.3 million. Keep in mind that Hugging Face counts file fetches, not people, so these numbers are inflated. The most common derivatives are post-trains, meaning further-trained versions of the finished model for a specific purpose, and distillations, where a model learns to deliver the same result in far fewer compute steps.
| Derivative | Who | What changed | Open weights? |
|---|---|---|---|
| Utopai X | Utopai Studios (USA) | Post-training on production data from film projects | no, only inside the PAI platform |
| HeyGen Video 1 | HeyGen (USA) | Post-training, details not disclosed | no |
| H3 Max / H3 Max Turbo | fal (San Francisco) | Post-training plus inference optimization | no |
| Pika H3 Flex | Pika (USA) | Fine-tuning, distillation from 50 to 6 steps, LoRA adapters | yes |
| HyperFlow H3 / H3 Flash | Video Rebirth | 8-step LoRA; Flash with a transformer shrunk to about 14B | yes |
| P-Video-2-Pro | Pruna AI | Tuned for speed, max 768p | no |
| FastH3 | Hao AI Lab (UC San Diego) | Distillation to 4 steps, sparse attention | yes |
| Open-H3-Regenerate-2K | Nvidia & MIT | Open replacement for the closed 2K module | yes |
LoRA adapters are small add-on modules that sharpen a large model without retraining it from scratch. On top of that come community tools. Redis creator antirez wrote his own engine for Apple Silicon Macs, h3.c. According to BigGo Finance, a data center in Chengdu ported H3 to Huawei's Ascend 910A chips and used it to produce an animated series for around 40 yuan per minute.
Our count across the AA leaderboards: in text-to-video with audio, 6 of the top 14 entries trace back to H3; in image-to-video, it's 7 of the top 12. The comparison with Meta's Llama is hard to avoid: an open base model that an entire industry builds on. According to AA, Utopai X beats its own base model in seven of ten categories, most clearly in audio sync, camera work and physics. AA shows a MiniMax logo next to the derivatives. But the same logo also appears next to FLUX 3 Max, which according to Unite.ai and VentureBeat is based on Black Forest Labs' FLUX 3. AA offers no explanation for the logo.
Do users ever find out?
Section IV.2 of the H3 community license requires commercial products to display "MiniMax H3" prominently in their interface. Our spot check of public documentation paints a mixed picture. HeyGen says it openly in its developer docs and its launch post on X: "Built on MiniMax H3, post-trained by HeyGen." Pika and fal put "H3" right in the product name. Utopai names the base model in its own materials. On OpenRouter, where you can also book HeyGen Video 1, there's no mention of MiniMax at all. We couldn't verify from the sources whether the consumer apps themselves show a notice. It's also unclear which license the US companies operate under. The US is itself an excluded territory, so these companies need special permission. The only public statement on this is a single sentence in fal's press release: the H3 team had "worked closely with fal from day one."
What a second of video costs
AA converts the prices of the vendors' APIs to a per-minute rate, using the benchmark's settings. From that, we calculated per-second and 10-second prices:
| Model | Elo (T2V with audio) | $/second | 10-s clip |
|---|---|---|---|
| Wan 3.0 | 1,174 | 0.20 | $2.00 |
| Utopai X | 1,168 | no API, approx. 1.40* | approx. $14* |
| MiniMax H3 (768p) | 1,159 | 0.08 | $0.80 |
| Seedance 2.5 | 1,156 | 0.57 | $5.69 |
| Gemini Omni Flash 1.2 | 1,151 | 0.10 | $1.04 |
| HeyGen Video 1 | 1,149 | 0.04 | $0.40 |
| fal H3 Max | 1,143 | 0.16 | $1.60 |
| Pika H3 Flex | 1,124 | 0.035 | $0.35 |
| Veo 3.1 | 963 | 0.40 | $4.00 |
*Utopai bills in credits: 93 credits per second, and the entry plan costs $15 for 1,000 credits.
That makes HeyGen half the price of the original while trailing by just 10 Elo points. Under the Elo formula, that works out to a win rate of roughly 51 to 49, and the confidence intervals (the uncertainty ranges around each ranking) overlap substantially. Prices vary by source. HeyGen's own blog lists $0.03 per second at 768p. According to its press release, fal charged $0.08 at 768p for H3 Max after the launch discount; AA lists $0.16. Seedance 2.5 Flash, ByteDance's new budget version, costs around $0.15 per second at 720p, according to a calculation by provider AIHubMix. It can do 30-second clips, but tops out at 720p.
The EU clause: who can do what in Germany
The license only applies in the "Applicable Territory," meaning the whole world minus the EU, the UK, South Korea and the US. In the excluded regions, Section V.4 doesn't just prohibit running and modifying the weights. It also bans using or displaying the generated output. That covers the open derivatives too: Pika H3 Flex is explicitly released under the H3 license. In its license FAQ, MiniMax justifies the restriction with regulatory uncertainty, for example around personality rights and copyright, and with ongoing copyright lawsuits in the US:
"The current restriction means 'not yet,' not 'never.'"
There's an application form for the excluded countries. Experiences with it vary widely. A CompanionLink blogger reported in September that his approval came through in under ten minutes. In the official license thread on Hugging Face, however, many questions go without a visible answer. A German hobbyist asked 13 days ago whether private individuals can apply at all, since the form requires a company name and a business email address. A US company writes that it submitted the form on November 2 and November 11 and has heard nothing beyond an acknowledgment of receipt.
Then there's Comfy, which says it's the only official reseller of MiniMax licenses. Comfy charges from $5,000 a month for commercial local use ("Professional" tier, up to 10 users). If you render on Comfy Cloud, commercial rights are already included in the subscription, according to Comfy.
None of this affects the APIs. According to MiniMax, its own API is available worldwide because the company can enforce safeguards there itself. The US derivatives can also be booked via API as usual. With HeyGen on OpenRouter, though, there's no EU routing, so the compute happens outside the EU.
| Route | Allowed for users in Germany? | Cost |
|---|---|---|
| MiniMax API (H3) | yes, under API terms | $0.08/s (768p) |
| HeyGen, fal, Pika via API | yes | approx. $0.035–0.16/s |
| H3 or derivatives locally | only with written permission | free to apply, outcome uncertain |
| Local commercial use via Comfy | yes, with license | from $5,000/month |
| Wan 3.0 (Frankfurt region) | yes | approx. $0.17/s (1080p, promo) |
| LTX-2.5 locally | yes; free for companies under $10M revenue, according to HeyGen | your own hardware |
Consultancy Contiamo argues that a weekend of tinkering is no problem. The license text, however, makes no such exception. Technically, H3 also runs on smaller hardware: Comfy cites an RTX 3060 with offloading to system RAM, and users recommend 64 GB of RAM for that.
Our verdict
Our take: in video, H3 is currently what Llama once was for language models, just with tougher terms. Well-known brands like HeyGen, Pika and fal refine a Chinese model and sell it for less than the maker itself. That's legal and makes a strong case for open weights, but it isn't disclosed openly everywhere. For users in Germany, it adds up to a paradox: the "open" model is exactly the one you can't simply run on your own machine here. The closed US derivatives, meanwhile, are just a credit card away. If you want to generate video cheaply and on solid legal ground, HeyGen Video 1 or direct MiniMax API access are currently your best bets. If you need top quality with data processing inside the EU, Wan 3.0 in Frankfurt is an option. If you want to work locally, you'll either have to wait for an answer from MiniMax or fall back on weaker, freely licensed models like LTX-2.5, which trails by more than 200 points on the leaderboard. Whether MiniMax will lift the restriction remains open. The newly released image variant, H3-Image, got the same clause, which suggests it won't happen anytime soon.
Sources
- AI Times: Video-KI, Umschwung binnen eines Jahres aitimes.com
- Artificial Analysis: Text-to-Video Leaderboard artificialanalysis.ai
- Artificial Analysis: Image-to-Video Leaderboard artificialanalysis.ai
- Artificial Analysis: Video Editing Leaderboard artificialanalysis.ai
- Hugging Face: MiniMaxAI/MiniMax-H3 huggingface.co
- MiniMax H3 Community License huggingface.co
- MiniMax H3 Lizenz-Q&A huggingface.co
- Hugging Face Diskussion: Any License Question Ask here! huggingface.co
- MiniMax: H3 Open-Source-Ankündigung minimax.io
- MiniMax API: Preise Video platform.minimax.io
- BigGo Finance: H3-Ökosystem finance.biggo.com
- HeyGen Developer Docs: HeyGen Video developers.heygen.com
- HeyGen Blog: Cheapest Video Models heygen.com
- OpenRouter: HeyGen Video 1 openrouter.ai
- Pika: H3 Flex Research pika.art
- Hugging Face: Pika-Labs/Pika-H3-Flex huggingface.co
- fal: Introducing H3 Max fal.ai
- PR Newswire: fal launches H3 Max prnewswire.com
- AlphaSignal: Utopai X schlägt sein Basismodell alphasignal.ai
- OrcaRouter: HeyGen Video 1 Launch orcarouter.ai
- Comfy Blog: Offizieller Reseller von H3-Lizenzen blog.comfy.org
- Comfy Support: MiniMax commercial licensing support.comfy.org
- Comfy Blog: H3 Day-0 Support blog.comfy.org
- Hugging Face: Comfy-Org/MiniMax-H3 huggingface.co
- GitHub: antirez/h3.c github.com
- The Batch: MiniMax's model is only minimally open deeplearning.ai
- Contiamo: MiniMax H3 and the small print contiamo.com
- CompanionLink: How to use MiniMax H3 in US or EU companionlink.com
- Atlas Cloud: H3 Commercial Use License atlascloud.ai
- Alibaba Cloud Model Studio: Preise alibabacloud.com
- dev.to/AIHubMix: Seedance 2.5 Flash vs 2.5 und 2.0 dev.to
- DataNorth: MiniMax H3-Image Lizenz datanorth.ai
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.