Keston CollinsVideo editor with nearly 10 years of experience, exploring the intersection of motion graphics and AI.
MiniMax H3 is the official name of the same model that Runway labels Hailuo 3.0. If you are weighing minimax h3 vs hailuo 3.0, there is no model choice to make. The difference is the label used by each app, API provider, or weights repository, including the shorthand Hailuo 03.
The naming split became visible after H3 was released publicly on July 31, 2026. MiniMax calls it H3 in its research material, API, and weights repository, while several product interfaces use a Hailuo label. The front-page results checked on August 17 agreed that the names refer to one model, but none assembled the platform labels into one map.
MiniMax H3 vs Hailuo 3.0: the platform name map
Here is the practical translation table. Use the string in the second column when searching a model selector or configuring that provider.
Platform or source
Name or identifier shown
What to do
MiniMax research blog
MiniMax H3
Use this name when discussing the official model
MiniMax platform API
MiniMax-H3
Use this model ID on MiniMax's API
Hugging Face
MiniMaxAI/MiniMax-H3
Use this repository name for the downloadable weights
Runway
Hailuo 3.0
Search this exact label in the model dropdown
fal image-to-video
minimax/h3/image-to-video
Use this endpoint for image-to-video requests
fal text-to-video
minimax/h3/text-to-video
Use this endpoint for text-to-video requests
fal reference-to-video
minimax/h3/reference-to-video
Use this endpoint for reference-to-video requests
OpenRouter
URL path minimax/hailuo-3, heading MiniMax: H3
Expect both labels on the same page
Vercel AI Gateway
minimax/minimax-h3
Use this provider-specific model ID
Replicate
Not available
Its newest listed MiniMax video models are Hailuo 2.3 and Hailuo 2.3 Fast
The unusual case is OpenRouter. Its URL uses hailuo-3, while the page heading says MiniMax: H3, without explaining the difference. That is still one model listing, not two choices.
MiniMax's own H3 research announcement does not use the phrase “Hailuo 3.0.” It calls the two earlier generations Hailuo 01 and Hailuo 02, then introduces the new model as MiniMax H3. This explains why a search across official and third-party interfaces can appear inconsistent even when the underlying choice is the same.
H3, M3, and Hailuo 2.3
H3 is not the video edition of M3. The materials identify M3 as a separate text and agentic model, while H3 is a general-purpose multimodal generation model that works across text, images, video, and audio.
Hailuo 2.3 is also not another name for H3. It is a different MiniMax video model and remains the newest MiniMax generation available on Replicate as of August 17, 2026.
The useful shortcut is simple: H3, MiniMax-H3, Hailuo 3.0, and Hailuo 03 point to H3 in the contexts mapped above. M3 and Hailuo 2.3 do not.
For generated-video results rather than naming, see the MiniMax H3 examples. If the model itself is unavailable on a chosen service, the MiniMax H3 alternatives cover other options.
Using the name in a real workflow
In Runway, open the model dropdown at the bottom of the panel and search for Hailuo 3.0. The Runway help center lists it under that name in Custom and Workflows on paid plans. It describes the provider as MiniMax but does not display the model name MiniMax H3.
For fal, copy the endpoint matching the input mode. The documented image-to-video call uses:
const result = await fal.subscribe("minimax/h3/image-to-video", {
input: {
prompt: "The camera slowly pulls back from the scene."
}
});
The endpoint name contains h3, not hailuo. In my workflow, I would copy the provider's exact identifier from the table instead of trying to normalize every service to MiniMax-H3. These strings are routing identifiers, so the official model name is not automatically the correct API value everywhere.
Open weights and hosted model
The downloadable MiniMax H3 repository is part of the official H3 system, but it does not reproduce every hosted component by itself.
The complete system has three modules. H3-Base generates audio and video at 768p. H3-Context-IR interprets free-form multimodal inputs and converts their relationships into a structured representation. H3-Regenerate-2K feeds the 768p result and original context back through H3 for a 2K result.
H3-Context-IR is not included in the open-source release because it relies on a hosted, multi-stage workflow and additional models and services. H3-Regenerate-2K is the module that lifts a result to 2K. The initial release also provides full-attention inference, while sparse-attention inference is planned for a later release.
The part nobody tells you is that downloading the weights does not give you the full official hosted pipeline. It gives you the released model components, while the hosted preprocessing and orchestration layer remains separate.
Practically, that means a local run and a hosted run are not guaranteed to be the same thing, even though both are correctly called H3.
After the clip
H3 returns a clip. Turning that clip into a repeatable branded format, with editable text and a layout you can re-render for the next launch, is a different job from generating it.
AutoAE is a video creation platform that covers that second step in the browser: pick a template, swap the text, logo, and media, then render. One example of that format:
Yes. MiniMax H3 is the official model name, while Hailuo 3.0 is the label used in Runway and associated with the Hailuo product experience. Hailuo 03 also refers to H3. The platform map matters because each API provider may expose the same model under a different identifier.
What is the relationship between H3 and M3? Is H3 the video version?
They are separate models. M3 is identified as a text and agentic model. H3 is a general-purpose multimodal generation model that understands context across text, images, video, and audio. H3 is not a video edition of M3.
How do I find it in Runway?
Open the model dropdown at the bottom of the panel and search for Hailuo 3.0. Runway lists it under that name in Custom and Workflows on paid plans. Searching only for MiniMax H3 may not match the label shown there.
Which minimax h3 model id should I use in an API?
Use the identifier required by the provider. MiniMax uses MiniMax-H3. fal uses endpoints such as minimax/h3/image-to-video. OpenRouter uses minimax/hailuo-3 in its URL, and Vercel AI Gateway lists minimax/minimax-h3. There is no single portable model ID shared by every service.
Are the Hugging Face weights the same as the model in the app?
They are official MiniMax H3 weights, but the download does not contain the entire hosted system. H3-Context-IR, the preprocessing and orchestration module used to interpret complex multimodal instructions, is not included in the open-source release. The repository provides an API route for reproducing the official workflow and guidance for building a separate preprocessing system.
