MiniMax H3, Hailuo, H3 Max — which name is which

Four of the names are not alternatives to each other at all. They sit at different levels: a company, a product line, a model, and the endpoints that serve it.

So “is Hailuo 3 the same as MiniMax H3?” — yes. “Is H3 Max just H3 renamed?” — no: it is a separate, post-trained endpoint with different limits, which is the one distinction in this list that actually changes what you get.

Three endpoints, and the honest trade between them is speed and resolution. Prices are per the model catalogue and update with it.

The audio column is the one that surprises people, and it is worth being precise about. The H3 model generates audio natively — the official card specifies 32 kHz stereo. Whether it reaches you depends on the endpoint: the Max endpoints return synchronised audio on every render with no toggle, while the base H3 endpoint we serve returns video only. The capability is the model’s; the behaviour is the endpoint’s. Those are two different things, and conflating them is why specs you read elsewhere can look wrong when you try them.

The official release ships distinct checkpoints for distinct input shapes, and their names leak into community discussion as though they were separate products.

H3-Base-FL2VA

First-and-last-frame mode. Takes zero, one or two images, so it covers text-to-video, first-frame, last-frame and first-and-last-frame in one checkpoint. If you have seen people refer to “fl2va” as though it were a separate model, this is it — it is a mode name, not a different model.

H3-Base-Ref2VA

Omni-reference mode. Accepts up to 9 images, plus up to 3 clips of video or audio at 2–15 seconds each. This is the checkpoint behind reference-to-video.

H3-Regenerate-2K

The upscaling module that takes output to 2K. The official model card lists it as not yet open-sourced — which matters more than it sounds, see below.

H3 is open-weights, so running it yourself is a real option, and plenty of people do. One detail is worth knowing before you spend an evening on it: 2K output comes from H3-Regenerate-2K, a separate upscaling module, and the official model card lists that module as not yet open-sourced. The base weights default to a shorter side of 768 pixels.

That is the gap between “H3 does 2K” in a spec sheet and what a local checkpoint actually produces. A hosted endpoint can offer 2K because the provider runs the whole pipeline; the downloaded weights on their own do not include the last step.

The licence is also worth a glance rather than an assumption: the release is under the MiniMax H3 Community License Agreement, not a blanket permissive licence.

These come from platforms and from the community, not from MiniMax. We do not serve any of them, so what follows is what each name refers to — and nothing about performance, because we have not measured them.

H3 Fast

A platform tier, not a MiniMax release. Some providers publish a cheaper, faster H3 endpoint under this name; what it actually maps to is that provider’s business, so check their own model page rather than assuming it matches anything here.

FastH3

A community step-distilled build of the open weights, aimed at cutting the number of sampling steps. Not from MiniMax, and not something we serve.

VDN-H3

A community port applying a different attention/decoding scheme to the open weights, circulated mainly among people running H3 locally. Again: community work, not an official variant.

H3 Context

A workflow for extending or continuing a clip, discussed mostly in local-generation circles. It is a technique built on top of the weights rather than a model you pick.

A general rule that saves time: when a platform advertises an H3 variant, the useful question is not “is this the real H3” but “which endpoint, at what resolution ceiling, at what price”. Providers pick those independently, so the same model name can behave and cost quite differently depending on where you run it.

Are MiniMax H3 and Hailuo 3 the same model?

Yes. MiniMax is the lab, Hailuo is its consumer product line, and H3 is the model both names point at. Writing it as “Hailuo 3.0” or “MiniMax Hailuo H3” does not change what is running underneath.

What is the difference between MiniMax H3 and MiniMax H3 Max?

Base H3 is the open-weights model, and on our catalogue it renders at 2K with reference-to-video support, priced per second at a flat rate. H3 Max and H3 Max Turbo are real-time endpoints post-trained on it: much faster and much cheaper per clip, with always-on audio, but capped at 768p and without a reference mode. Speed and resolution are the trade.

What is the difference between H3 Max and H3 Max Turbo?

Same capabilities, different speed and price point. Turbo renders a 5-second 768p clip in roughly 1.5 seconds upstream at half the per-second cost; the standard Max endpoint takes about 2.7 seconds and costs twice as much. Turbo is the default in Live Mode; Max is the “higher quality” toggle.

Can I run H3 on my own GPU?

The base weights are published, so yes — that is why you see people fitting it onto consumer cards and trading quantisation notes. Two caveats worth knowing before you start: the licence is the MiniMax H3 Community License Agreement rather than a blanket permissive licence, and the 2K upscaler is a separate module that the model card lists as not yet open-sourced.

Why do local runs not reach 2K?

Because 2K comes from H3-Regenerate-2K, a separate upscaling module, and the official model card marks it as not yet open-sourced. The base weights default to a shorter side of 768 pixels. A hosted endpoint can offer 2K because the provider runs the full pipeline; a local checkpoint on its own does not.

Does H3 generate audio?

The model itself does — the official card specifies 32 kHz stereo. Whether you get that audio depends on the endpoint you are calling. On this site, H3 Max and H3 Max Turbo return synchronised audio on every render with no toggle, and the base H3 endpoint we serve does not return an audio track. That mismatch between model capability and endpoint behaviour is the single most common reason people think the specs they read were wrong.

Which one should I use here?

Want it fast and cheap, and 768p is enough: H3 Max Turbo. Want the same thing with more headroom on quality: H3 Max. Need 2K, or need reference images to hold a character or product: base MiniMax H3. Need longer than 15 seconds: none of them — look at Seedance 2.5, which runs to 30.

Official capabilities, mode names and licence last verified against MiniMax’s model card on September 8, 2026. Endpoint specs and prices come from this site’s model catalogue and change with it.