MiniMax just launched an AI video model that listens to a photo, watches a video clip, and reads an audio file in the same prompt, then hands back a 2K video with its own stereo soundtrack built in. It also open sourced part of the model days later, then quietly excluded the US, the EU, the UK, and South Korea from using those weights locally. Here is what H3 actually does, how to try it, and the access catch most coverage is skipping past.
Most AI video tools still make you pick a lane. Text to video. Image to video. Upload a reference clip for motion, upload a separate file for voice, stitch the outputs together in an editor afterward. MiniMax built H3 to skip that entirely. You describe the relationship between your inputs in one sentence, and the model figures out the rest.
That alone would be a solid product update. What makes H3 worth a longer look is everything that happened in the week after launch: an aggressive price position against Kling and Seedance, a benchmark win, a stock pop, an open weights release that turned out to be partially open, and an active Hollywood copyright lawsuit sitting underneath the whole platform. None of that shows up in a two line product announcement.
What MiniMax Actually Launched
MiniMax announced H3 on July 31, 2026, after teasing it under the #MiniMaxH3 tag the day before. The pitch: one omni modal model that jointly understands text, images, video, and audio as a single context, then generates video with native stereo sound at up to 2K resolution and 15 seconds in length.
The demo MiniMax showed off is the clearest way to understand what is actually new here. The prompt was roughly: reference the camera movement from this video clip, have the character in this photo sing, match the vocals to this audio file. H3 took all three inputs and produced one coherent video. No separate motion reference tool, no separate lip sync pass, no manual audio mixing afterward.
MiniMax is positioning H3 for commercial content work specifically: advertising, branding, e-commerce, product design, UI and UX mockups, gaming assets, and film opening titles. Early testing highlighted instruction following, accurate on screen text and brand rendering, and video to video motion transfer as particular strengths.
On pricing, MiniMax’s own claim is that 2K generation costs less than a third of what mainstream competitors charge, and 768p generation costs less than half of what competitors charge at 720p. The company did not name which models it was comparing against. One independent breakdown puts H3’s 2K API pricing at roughly $7.80 per minute, against roughly $20 to $22 per minute for Kling 3.0 and Seedance 2.0 at 1080p, which lines up with that claim reasonably well. More on that comparison below.
How to Actually Try H3 (Step by Step)
You do not need a developer account to see what H3 can do. Here is the fastest path in.
- Open the Hailuo AI demo tool. MiniMax put H3 directly into its consumer app at hailuoai.video. H3 has replaced Hailuo 2.3 as the app’s current default model, so you may already be generating on it without realizing.
- Pick your inputs. You can start from text alone, or add a reference image, a reference video clip, and a reference audio file in the same generation. This is the actual point of H3, so it is worth testing with at least two input types before judging the model on a plain text prompt.
- Write the relationship, not just the scene. Instead of describing what you want to see, describe how your inputs relate to each other. “Use the camera movement from this clip, apply it to this character, match this voice” works better than a generic scene description, because that relational instruction is exactly what H3 was trained to parse.
- Generate at 768p before you pay for 2K. 768p renders are cheaper and faster, and they are the right place to iterate on a prompt before committing to a full resolution render.
- For production work, use the API. If you are building H3 into a workflow rather than generating one off clips, MiniMax’s platform API and third party hosts including Segmind and fal both have H3 endpoints live.
The Open Weights Come With a Catch
This is the part of the H3 story that most first week coverage missed, because it happened three days after launch and outside the original announcement.
MiniMax published H3’s weights to Hugging Face on August 3, 2026, under a license MiniMax calls the H3 Community License. Reasonable enough, until you read what it actually grants. The license defines an “Applicable Territory” for local deployment, and that territory specifically excludes the United States, the European Union, the United Kingdom, and South Korea. Users in those regions are not licensed to run, modify, or deploy outputs from locally hosted H3 weights.
MiniMax’s Head of Developer Relations, Ryan Lee, confirmed publicly that the US exclusion ties directly to the company’s active copyright litigation, which we will get to in a moment. MiniMax’s own license documentation adds a second reason: video models sit in a messier regulatory spot than text or code models right now, with the EU AI Act, evolving UK and South Korean rules, and an unsettled US landscape all pointing the same direction. Once weights are public, the company cannot enforce safeguards downstream, so restricting the license became the alternative to delaying release.
There is a second layer worth knowing before you plan a self hosted deployment around H3. What actually shipped on August 3 is H3-Base, a 33.1 billion parameter model that generates natively at 768p. The 2K output featured in every promotional clip comes from a separate module, H3-Regenerate-2K, which was not open sourced and remains an API only feature. So even a fully licensed, self hosted H3 setup still needs to call MiniMax’s servers to get the resolution the marketing is built around.
If you are outside the excluded territories, or MiniMax approves an individual license for your use case, the Community License is free for organizations under $20 million in annual revenue, on the condition that “MiniMax H3” appears in your product’s interface. It also bans using H3 outputs to train or improve a competing model, a restriction that applies everywhere, license region or not.
How H3 Stacks Up on Price
Independent pricing comparisons for AI video models are still thin, since most vendors quote per second or per credit rather than a clean per minute rate. The most direct side by side comparison available puts H3 meaningfully ahead of two of its closest closed source competitors at comparable resolutions.
| Model | Resolution compared | Approx. cost per minute | Native audio |
|---|---|---|---|
| MiniMax H3 | 2K | ~$7.80 | Yes, native stereo |
| Kling 3.0 | 1080p | ~$20.16 | Not confirmed |
| ByteDance Seedance 2.0 | 1080p | ~$22.45 | Not confirmed |
Worth flagging: this specific comparison comes from one independent explainer, not a controlled MiniMax or third party benchmark, so treat the exact dollar figures as directional. What is more solidly confirmed is where H3 sits on quality. Artificial Analysis, an independent benchmarking platform, ranked H3 first globally in video editing at launch and placed it in the top three for both text to video and image to video. It trailed Google’s Gemini Omni Flash on text to video, and trailed both Gemini Omni Flash and ByteDance’s Seedance 2.0 on image to video specifically. First in editing, competitive but not dominant everywhere else, roughly what you would expect from a model built around reference and editing tasks first.
The Lawsuit Sitting Underneath Every Clip
Here is the part that changes how you should think about using H3 for anything client facing or commercial. Disney, Universal, and Warner Bros. Discovery jointly sued MiniMax and its Hailuo platform in September 2025, alleging the company trained its video and image models on unauthorized copies of their characters, and that Hailuo could generate recognizable figures including Spider-Man, Darth Vader, and Shrek from a simple text prompt.
MiniMax tried to get the case dismissed on jurisdictional grounds, arguing US courts had no authority over a Chinese company. On May 26, 2026, a federal judge rejected that argument and let the case proceed into full discovery. That ruling does not mean MiniMax did anything wrong. It means the studios presented enough of a case to keep going. But it does mean the platform H3 is built on is currently under active US litigation over exactly the kind of output generation H3 is designed to do well.
H3 runs on the same Hailuo infrastructure named in that lawsuit. If your use case is internal experimentation, this is background context. If you are generating client work, advertising, or anything with commercial IP exposure, it is a live legal question worth flagging to whoever signs off on your content pipeline.
Why MiniMax Is Doing This Now
None of this is happening in a vacuum. ByteDance rolled out its own upgraded Seedance 2.5 the same week H3 launched, and the Hong Kong market reacted to MiniMax’s move immediately, with shares climbing roughly 13 percent intraday on heavy turnover. Video generation has been one of the last corners of generative AI where closed, expensive models still had a comfortable lead. Undercutting on price and partially opening the weights is MiniMax spending that lead deliberately, betting that distribution and developer adoption matter more right now than keeping every advantage closed.
It is a familiar playbook by now. Ship something genuinely capable, price it well under the incumbents, open just enough of it to build goodwill and an ecosystem around it. Whether that trade pays off depends on whether H3’s actual output quality holds up once the initial wave of testing settles down, and whether the license restrictions loosen once the copyright case moves further along.
Limitations to Keep in Mind
- 15 seconds is a hard ceiling per clip. Anything longer needs to be stitched from multiple generations, and Seedance 2.5’s 30 second clips are a real advantage for longer narrative work.
- 2K output requires an API call to MiniMax’s servers even if you are running the open weights locally, since the regeneration module that produces 2K was not open sourced.
- If you are in the US, EU, UK, or South Korea, self hosting H3 legally requires an individual license from MiniMax. The default Community License does not cover you.
- Every API call, regardless of where you are, routes through infrastructure in China, which is subject to China’s National Intelligence Law. That law can compel Chinese organizations to cooperate with government data requests regardless of server location.
- MiniMax’s pricing claims and use case demos are the company’s own promotional material. The Artificial Analysis benchmark scores are independent, but reflect the hosted API pipeline, not a locally run H3-Base render.
- No independent security audit of the released weights exists yet, and training data provenance has not been independently verified.
Do This Today
Right now: Head to Hailuo AI and run one test generation using at least two reference inputs, an image plus an audio clip, or a video plus text, rather than a plain text prompt. That is where H3 actually differentiates itself from a standard text to video tool.
This week: If you are evaluating H3 for client or commercial work, read the licensing question before the creative question. Confirm which territory you are operating from and whether the Community License actually covers your use case, or whether you would be relying on the API instead.
This month: Price out a real project against H3, Kling 3.0, and Seedance, using your actual clip lengths and resolution needs rather than the headline per second rate. The cost advantage is real, but it only matters if the output quality clears your bar for the specific kind of content you make.
An AI model that listens to a photo, a video clip, and a voice recording all at once is genuinely new. A model that ships “open weights” you are not allowed to run in four of the world’s biggest developer markets is a reminder to always read past the headline.
Sources
- MiniMax, MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities, July 31, 2026
- South China Morning Post, Video AI: MiniMax challenges ByteDance with low price, open weights for new H3 model
- Tech Times, MiniMax H3 Open Weights Exclude US, EU, UK, and Korea From Local Deployment
- Tech Times, MiniMax H3 Opens AI Video to Developers: Copyright Lawsuit Clouds Every Clip
- Crypto Briefing, MiniMax loses bid to end Disney copyright lawsuit over AI system
- Hugging Face, MiniMax H3 model card and license
- Segmind, MiniMax H3: Release Date, Open Weights, and API Pricing Explained
- Artificial Analysis, Video Generation Leaderboard
Just Being Resourceful! 
Leave a comment