MiniMax M3 reads video. It cannot render a single frame — here is the model that can.

MiniMax M3 does not generate video, and never has. It takes video in — up to around thirty minutes of footage in one request — and writes text out. Descriptions, critiques, shot lists, code. The MiniMax model that renders video is H3, released two months after M3, one letter away in the API string.

There are two ways you got here and they need opposite answers.

If you came looking for a MiniMax M3 video generator, the model you want is called H3 and you can make a clip with it now — same company, right letter. If you actually meant M3 and want to know what it does with footage, keep reading. The answer is a lot more specific than "nothing", and the real limit is not the one on the spec sheet.

Can MiniMax M3 generate video?

No. I went to the pipeline tag first, because it settles the question in one lookup and you can repeat it in under a minute.

The pipeline tag is the fast one. It is a declared field on the model repository and it names the output modality — M3's says text. A model whose task tag ends in -to-text has no path to a video file, however the request is worded.

The billing row is the one that bites teams rather than individuals. Price a project on the wrong model here and the estimate is not slightly off, it is in the wrong unit — tokens against output seconds — and nothing downstream reconciles it.

Feed M3 a clip and here is what it can actually see

Worth knowing even if you only ever make video, because M3 turns out to be useful on the way in.

M3 is natively multimodal: text, images and video all arrive through the same chat endpoint, trained together from the start rather than through a vision adapter bolted onto a text model. NVIDIA's deployment card puts the long-form ceiling at about 30 minutes of footage per request, and MiniMax reports 84.6 on Video-MME at 512 frames — its own evaluation, so read it as vendor-reported.

Three things matter more than that score.

Video goes in as frames, not as a file. You sample the clip and pass an ordered sequence of images. MiniMax's own published evaluation sampled at 1 FPS.

Which makes the ceiling a token budget wearing a duration costume. Every frame you sample spends part of the 1,048,576-token context window. That is why the limit appears as minutes on a deployment page and as frames in an evaluation config: 512 frames at 1 FPS is about eight and a half minutes of real footage. Half an hour is half an hour only if you are willing to look at it one frame per second — sample two minutes at 12 FPS and you have spent the same budget on a twelfth of the runtime. Set your frame rate against the shot, not against the file length.

M3 cannot hear your clip. The video path is a sequence of stills and the soundtrack does not travel with them. The tell is in MiniMax's own Video-MME setup, which interleaved subtitles as text every 30 seconds instead of feeding audio in. So anything M3 says about dialogue, music or room tone, it inferred from pictures and whatever text you handed it. If the dialogue matters, paste a transcript.

Which leaves a neat inversion to remember the pair by: H3 is defined by the sound it produces, M3 by the sound it cannot take in. Same company, two months apart.

If describing footage you already have is the real job, our video to prompt tool does the H3-shaped version: it reads a clip and returns a structured prompt you can render straight from, instead of prose you then have to convert by hand.

Four names now answer to "MiniMax 3"

Three products with a 3 in the name shipped in 2026, the API strings for two of them differ by one character, and MiniMax sells M3's million-token window on "long-video understanding" — accurate, and indistinguishable from video generation at a glance.

H3 Max is the newest way to land in the wrong place. It is a real MiniMax-sanctioned model, listed beside H3 in MiniMax's own documentation, but not one you can swap in freely: hosted only, no published weights, 768p rather than 2K. What "Max" actually maxes, and what it gives up walks the whole trade.

Which model your job needs

Let M3 write the prompt, and H3 render it

M3 does have a job in a video pipeline. It is just not the rendering job.

H3 does not want a sentence. It wants a long structured description — shot by shot, with camera moves, timings and sound written into named fields. MiniMax builds that description with a hosted stage called Context-IR that is not in the open release, and its own guidance invites developers to build their own preprocessing instead.

M3 fits that opening almost too well. A million-token window holds a brand style guide and a shot list at the same time; native image and video input means it can read your reference material; and it outputs text, which is exactly what H3 consumes. So the pipeline is M3 writes the prompt, H3 renders it — and the search that brought you here was one letter away from a real workflow rather than a mistake.

The full handoff — the system prompt that makes M3 emit H3's three-field format, what the exchange costs against the render, and where the two licences diverge — is in MiniMax H3 vs MiniMax M3. To skip building it, our prompt generator produces the same structure in the browser.

MiniMax M3 and video: FAQ

Is MiniMax M3 a video model? It is a multimodal model that reads video. It is not a video generation model — its pipeline tag is image-text-to-text.

Can MiniMax M3 do text to video? No. H3 does, and it is the model MiniMax built for it.

What is the "MiniMax M3 video generator" I keep seeing advertised? Almost always a MiniMax H3 interface using the M3 spelling to catch this search. The generator is real; the model name on it is wrong.

Is H3 Max the same thing as M3? No, and these two are the easiest pair to swap by accident. H3 Max makes video and is a post-trained H3; M3 makes text. The shared letter is a coincidence.

Does M3 understand the audio in a video I upload? No. Frames go in, the soundtrack does not. Paste a transcript as text when the dialogue matters.

How many parameters does MiniMax M3 have? About 428B total, roughly 23B active per token, across 128 experts with four activated per token. Hugging Face's safetensors metadata reports 427,040,140,160 exactly.

Are MiniMax M3 and MiniMax H3 both open weights? Both publish weights under restricted licences that are not OSI open source, and the restrictions differ in shape — M3 carries no territory limit, H3 excludes four markets. What the H3 licence actually restricts breaks that side down.

I just want to make a video. Start with MiniMax H3 text to video — no M3 required, and the sound arrives in the same pass.

---

Last verified 31 August 2026. The 30-minute ceiling is NVIDIA's published deployment figure rather than a MiniMax API guarantee, and the Video-MME score is MiniMax's own — both are sourced so you can re-check them rather than take them on trust. MiniMax ships fast enough that a fourth "3" is a live possibility.