How to use MiniMax H3: five routes into the same model

There is no single MiniMax H3 app. It is a model behind five doors — this browser, MiniMax's own apps, the official API, a reseller, or the weights on your own GPU — so how you use it depends entirely on which one you walk through.

All five run the same weights and return a 4-to-15-second MP4 at 24 fps, 768P or 2K, with sound alongside the picture. What changes is setup time, who pays, and whether you are licensed at all: the hosted routes are open worldwide, while running the weights yourself is not licensed in the US, the EU, the UK or South Korea.

Five ways to get started, ranked by setup time

Sorted by work required before your first clip, not by the row we would prefer.

The last column matters most. MiniMax's community licence puts the US, the EU, the UK and South Korea outside its Applicable Territory, and the exclusion follows the outputs too. Hosted routes are untouched. So from London or San Francisco, rows one to four are yours and row five is not — which is most of the answer to how to use MiniMax H3. A summary, not legal advice.

Route one: in a browser, in about 90 seconds

Type a shot, press generate. No account, no email, no card — a bot check, then a 768P clip you can download and keep, unwatermarked.

Your text does not go straight to the raw weights. It goes through MiniMax's hosted pipeline, which runs the prompt-understanding and 2K stages the download does not include. Text to video and image to video expose every parameter.

Route two: MiniMax's own apps

Which pair you get depends on where you are. Globally, Hailuo AI at hailuoai.video and a desktop client at hub.minimax.io; in mainland China, hailuoai.com and hub.minimaxi.com. Both want an account and bill in MiniMax's own credits.

A naming trap waits there. Hailuo 3 is this model — MiniMax's own OpenAPI title reads Hailuo-03. Hailuo AI without a number is the app, and Hailuo 2.3 is the previous generation.

Route three: the official API

Get a key at platform.minimax.io, or platform.minimaxi.com for a China account, and point at the matching base URL. The model ID is MiniMax-H3, case-sensitive. The console builds the call; the rest is what no snippet teaches.

Nothing here is synchronous

The create call returns a task_id, not a video. Poll GET /v2/query/video_generation/{task_id} until the status reads succeeded, then read the file from content.url. To skip polling, set callback_url: MiniMax first sends a verification request with a challenge field you must echo back within three seconds.

Two clocks start the moment it succeeds

Two separate clocks, and almost everyone merges them. The output URL is time-limited, so download and re-host on completion; MiniMax publishes no figure for it, and the 24-hour number circulating elsewhere is a reseller's. The task stays queryable for seven days, after which the same task_id returns invalid.

Five things that break the first call

The one that catches almost everyone is ratio — required for text-to-video, rejected if you set it to adaptive. A content array with no non-empty text item returns error 2013.

The one that costs money is the last: a reference request with nothing attached does not fail. It becomes a text-to-video job, returns 200 OK, and bills you.

Route four: resellers, with a different vocabulary

Several platforms host H3 behind their own schemas, with their own field names and their own limits. Code from a reseller's docs will not run against api.minimax.io: quality is resolution here, aspect_ratio is ratio, and the model ID is not hyphenated. Some accept 2K only, which is where that claim comes from; MiniMax's own API takes both.

Context-IR and Regenerate-2K are not in the download

H3 runs in three stages and only the middle one is open.

• H3-Context-IR rewrites your prompt and reference material into a structured shot description.
• H3-Base generates the 768P picture and its audio in one pass.
• H3-Regenerate-2K runs that result back through the model with the original context, which is why small text survives instead of being guessed at.

The weights on Hugging Face are H3-Base. The other two stayed hosted. That is why a local render looks flatter than the launch reel with your settings perfectly right — you are running one stage of three.

Call the hosted API and all three run for you. Less known: the two closed stages are callable on their own. MiniMax's own model card documents a hybrid — H3-Base on your SGLang deployment plus the official Context-IR and Regenerate-2K APIs — as the way to reproduce the 2K the hosted API gives you directly. Self-hosting is not all-or-nothing.

Route five: your own GPU, and when not to use us

Only if the licence covers you — the table above says where it does not. It reaches the outputs too: render in a permitted country, publish in an excluded one, and you are still outside the grant. MiniMax takes applications at platform.minimax.io/h3-license. The licence, clause by clause has the rest. A summary, not legal advice.

If you are clear: the ComfyUI compact package is 42.47 GB, not the 498 GB of the full repo, and ComfyUI v0.30.0 or later reads it without custom nodes. Running it in ComfyUI has the file layout; when a GPU pays for itself has the arithmetic.

Three cases where you should not use this site. Commercial work: go to MiniMax's own Hailuo product or the API, because everything made here is for personal and evaluation use. Volume, or H3 inside your own product: that is the API. Files you need next month: keep your own copies; we are not an archive.

The questions people ask first

Is there a free way to use MiniMax H3? Yes, and you are one click from it. Route one gives a 768P five-second clip with no account and no card, behind a bot check. Signing in adds credits. Nobody offers unlimited free H3: every output second is real GPU time, billed at $0.08 on the API.

What is MiniMax H3? MiniMax's omni-modal generation system. It reads text, images, video and audio as one context and returns 4 to 15 seconds of 24 fps video at 768P or 2K, with 32 kHz stereo audio in the same pass.

When was it released? Announced 31 July 2026; open weights published 2–3 August 2026. The community licence is dated 2 August.

Is MiniMax H3 the same as Hailuo? Partly. Hailuo 3 is the same model. Hailuo AI without a number is the consumer app, and Hailuo 2.3 is the previous model.

What is the API model ID? MiniMax-H3, case-sensitive. Lowercase hyphenated IDs belong to resellers, not to this API.

Why is my text-to-video call failing? Most often ratio, which text-to-video requires and which rejects adaptive. Next is a content array with no non-empty text item, which returns error 2013.

What does it cost? $0.08 per output second at 768P and $0.13 at 2K on MiniMax's pay-as-you-go list. Reference video is billed by its own length on top; the first five reference images and all reference audio are free. 768P costs 38% less, so drafts belong there.

Can I run it on my own machine? Only where the licence allows, and only H3-Base. The other two stages stay hosted, but a local deployment can still call them.

Can I edit a clip after it is generated? Not with an edit instruction. H3 has no edit endpoint; a guide describing one is describing something else. Video goes back in through the reference path, and a new clip is generated from it.

Can I use what I make here commercially, and is it watermarked? Nothing we hand back is watermarked. Commercial use is the one no: everything generated here is for personal and evaluation use, on every plan including the free one. We are an independent third-party interface and we are not in a position to grant you commercial rights to the output. For commercial work, generate through MiniMax's own Hailuo product or the official API, both of which come with commercial terms. A summary, not legal advice.

---

The shortest answer to how to use MiniMax H3 is the generator. No card, no account, and what you download is yours.