MiniMax ships its video generator under the name Hailuo, and the Chinese landing page puts the pitch in one line: every idea deserves to be a blockbuster. Two models sit at the centre of the product right now, H3 and H3 Max, announced in the same banner that offers newcomers 3000 free credits. The workspace itself stays narrow — video on H3, still images, audio — with three doors labelled create video, create image and Hailuo Agent.

Four modalities, one model

The positioning is a single sentence: an open-weight, general-purpose, omni-modal generation model, and the research index carries an Open Weights badge right next to it. The post that introduced H3 spends its opening on why the previous two generations looked the way they did. Hailuo 01 built the system from nothing. Hailuo 02 went after architecture efficiency, data quality and scale. H3 takes a different route and glues back together work that had been sliced apart.

On the image side that meant separate specialists for text-to-image, editing, subject reference, motion reference and style reference. Audio split into voice, sound effects and music, each studied on its own. Video fragmented further, into text-to-video, image-to-video, first and last frame, subject reference, motion reference, timbre reference and video editing. The borders between modalities were just as firm. Those walls limited what a creator could ask for and capped how far the model could generalise, so pulling tasks back under one roof became the first design rule.

What comes out: 768P or 2K, 15 seconds, stereo built in

Three claims open the post. H3 reads text, images, video and audio inside one context. It generates clips with native stereo sound. Footage runs up to 15 seconds at 2K, and 2K is the default rather than an upgrade. The documentation is more exact: H3 outputs 768P and 2K, with durations of 4 to 15 seconds in whole numbers only. H3 Max, released jointly with an AI inference platform and tuned for speed, outputs 480P and 768P at 5 to 15 seconds.

Audio carries much of the weight this generation. Every audio output is native stereo, and voice, effects and music are modelled together instead of being stitched on afterwards.

Three ways in

Text-to-video wants nothing but a prompt. Image-to-video accepts first-frame and last-frame pictures, and the count can be 0, 1 or 2 — pinning the opening frame, the closing frame, or both, in which case the model is left to fill the middle.

Reference generation mixes reference images, reference video and reference audio to lock a person, a motion, a camera move, a style, a timbre or an edit rhythm. The ceilings are spelled out: up to 9 images, up to 3 clips of 2 to 15 seconds each with no more than 15 seconds combined, up to 3 audio tracks, and 12 files in total for a mixed request.

2K arrives by regeneration, not by upscaling

Rather than bolt a dedicated upscaler onto the end of the pipeline, the team lets the base model regenerate its own low-resolution output inside the context — In-Context Regeneration. Two things follow. Generation ability the base model already has gets reused instead of duplicated in a second network, and the regeneration pass can call the original multimodal context again. That is how small lettering and fine texture, which a conventional upscaler can only guess at, come back.

There is a second door for footage you already hold. If a clip meets the 768P specification, a regeneration job can lift it to 2K. The request has to reproduce everything that produced the original, plus one extra entry for the source video itself.

Read the context before generating

Multimodal inputs get complicated quickly, and a hand-written prompt rarely captures them. H3-Context-IR is the official answer: it parses text, images, audio and video together, reasons about how the parts relate, and returns a richer structured description. It only hands back the enhanced prompt and produces no footage at all, which makes it a translation layer that runs before generation rather than a generator.

The mechanism behind it is Contextual Omni Representation, which uses language as a general-purpose bridge and interpreter and turns each task into an open description. The understanding pipeline distils roughly 100K tokens of raw material down to about 4K tokens on average before the model sees any of it.

The price claim

On value, the post says H3 gives 2K by default, costs less than a third of mainstream models per second at 2K, and less than half of what mainstream models charge for 720p when you run at 768P. That is the vendor's own framing, and readers can run their own numbers against it.

For individuals the Chinese site advertises from 42 yuan a month, the international site from 9 US dollars 99 cents a month, and both hand free credits to new accounts — the Chinese site names 3000 points.

Driving it from an API

Generation is asynchronous, and the documentation breaks it into three moves. Create a job and take the task_id. Poll that id; once the job succeeds the response carries the location of the finished clip directly, with no separate file id to exchange. Then fetch the file from that location. Polling every 10 seconds is the recommended rhythm — kinder to the servers — and the states failed and cancelled are terminal.

A few ceilings on the input side are worth knowing before the first call. Prompts run to 7000 characters. Single files run to 50 MB for video, 30 MB for images and 15 MB for audio, while request bodies run to 64 MB, with large assets better passed as a URL. Video arrives as H.264/AVC or H.265/HEVC, images as JPG, PNG, WEBP or HEIC, audio as WAV or MP3.

Where the fine print sits

Open weights remain a statement of intent. The site already labels H3 as open-weight, while the post says the weights will be opened later subject to applicable law and regulation. What actually ships is what counts, not what is planned.

More reference material is not automatically better. The 12 file ceiling on a mixed request is hard, and the Chinese and international sites show different defaults, so the real figure depends on the model and the tier selected.

The two models are not interchangeable. H3 handles 768P and 2K at 4 to 15 seconds; H3 Max handles 480P and 768P at 5 to 15 seconds. Match the model to what the finished clip needs before choosing.

The one-third figure is a vendor claim, and third-party prices and promotions move. Budget against the rate published at the time, and remember that generated material still has to clear platform rules and local law.

Summed up: the interesting part is not another text-to-video door. It is that H3 folds video, image and audio understanding into one context — 2K by default, up to 15 seconds, stereo from the start, 2K delivered by regeneration rather than upscaling, and reference budgets of 9 images, 3 clips and 3 audio tracks. Those numbers, more than the slogan, are what make it practical next to similar tools.