MiniMax H3 vs MiniMax M3: one makes video, one writes code

This compares two models that do not compete. H3 is an omni-modal video model: text, images, video and audio go in, and 4–15 seconds of video with sound in the same pass comes out. M3 is a 428B mixture-of-experts language model with a 1M-token context that reads video but writes text.

M3 cannot generate video. H3 cannot write code. If you are choosing between them, you have not yet decided what you are making — and if you are making video, M3 is still worth having, for reasons at the bottom of this page.

One makes video, one writes code

MiniMax H3 shipped on 31 July 2026, with open weights on 2–3 August. It takes text, images, video and audio, and returns a single MP4 at 24 fps, 4 to 15 seconds long, up to 2K. You call it at /v2/video_generation, asynchronously: create a task, poll the id, download the file. It bills per second of output.

MiniMax M3 shipped two months earlier, on 1 June 2026, with weights on 12 June. It is a 428B mixture of experts, 23B active per token, built on MiniMax Sparse Attention for a 1,048,576-token window. It reads images and video, and it writes text. You call it at /v1/chat/completions, synchronously. It bills per token.

Same host, same account, one letter apart in the model string.

Which one are you actually looking for?

Most people searching this are not comparing anything — they landed on the wrong model and need out. One question settles it.

What do you want to come out?

If you only have a name or a number and cannot tell which it belongs to, the two lookup sections below are for you.

The numbers side by side

Two rows decide almost everything.

What it outputs is the only row most readers need. Billing unit is the one that catches procurement: a quote built on the wrong model is not slightly wrong, it is priced in a different unit. H3 is $0.08 per output second at 768P and $0.13 at 2K. M3 is $0.30 per million input tokens and $1.20 per million output, doubling above 512K input.

One row is worth a second look. Both accept video — but H3 accepts it as a reference to imitate, and M3 accepts it to describe. Same input, opposite purpose.

The licences point opposite ways

Both are called a Community License. They are different documents, and they restrict different things.

The H3 licence excludes the European Union, the United Kingdom, South Korea and the United States from its Applicable Territory, and §V.4 extends that restriction to the outputs — a clip rendered legally elsewhere is not licensed for display in those four markets. Above $20M in annual revenue you need prior written authorisation.

The M3 licence has no territory restriction at all. Its base grant is non-commercial; commercial use requires you to display "Built with MiniMax M3" prominently, and the same $20M threshold triggers written authorisation, with a one-time notice below it.

So the two questions are different. H3 asks where you are. M3 asks how much you earn. A developer in London may legally self-host M3. They may not self-host H3.

One more contrast worth knowing: H3's text encoder is Qwen3-VL-32B, a third-party model, while M3's sparse attention is entirely in-house. The same company took opposite routes on its two flagships. And worth noting, M2 shipped under MIT in October 2025 — the restriction started at M2.7.

A summary, not legal advice. The full H3 licence breakdown has the clauses; if one of those four markets is yours, the hosted API is a separate arrangement.

Every name MiniMax has shipped

This is only part of a wider naming problem. MiniMax has shipped, in one year, a video model called H3, a language model called M3, and a music model called Music-3.0 — so "the new MiniMax 3 model" currently has three referents.

Writing Hailuo 03 is not a mistake; it is an extrapolation MiniMax invalidated. The video line has carried four naming schemes in four generations: T2V-01-Director, then MiniMax-Hailuo-02, then MiniMax-Hailuo-2.3, then MiniMax H3. Anyone following the 2025 pattern would predict Hailuo-03.

That guess then collides with Kling O3 — which is not a MiniMax product. It is Kuaishou's, the vendor calls it Video 3.0 Omni, and the two strings are nearly identical in a search box.

That number you saw — check which model it belongs to

A lot of the mix-up arrives as a figure. SWE-Bench Pro 59.0%, Terminal-Bench 66.0%, MCP Atlas 74.2% — those are M3's, they are vendor-reported, and no video model can be scored on them. If you see one next to H3, that page crossed its wires.

It runs the other way too. Elo ratings, per-second prices and the 42.47 GB ComfyUI download belong to H3, not M3. Both directions matter, because a wrong number in a procurement doc survives longer than a wrong name.

One figure is simply stale: M3's Artificial Analysis score of 55 predates Index v4.1, which now reads 44.

"vs" is the wrong question

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

M3 is unusually well shaped for that job, and it comes from the same vendor. It has a million-token window, so a brand style guide and a shot list fit inside it. It reads images and video, so it can derive a description from your reference material. And it outputs text, which is exactly what H3 consumes. This is not a workaround. It is the gap MiniMax left open, filled with MiniMax's own model.

So the pipeline is: M3 writes the prompt, H3 renders it, and M3 reviews the result and revises. The system prompt that makes M3 emit H3's format:

You are a MiniMax H3 prompt writer. Convert the user's idea into H3's structured
prompt format, following MiniMax's official h3-prompt-writing skill.

Rules:
- Output exactly three fields, in this order:
  integrated_multimodal_description, overall_soundscape, non_diegetic_music.
- Write the descriptive sections in English. Keep dialogue, lyrics and on-screen
  text in their original language.
- Break the clip into shots using [Shot 1], [Shot 2]. Mark cuts with timestamps
  like "At 00:04.500".
- Describe each shot in this order: composition, subjects, environment, actions,
  camera, sound.
- Describe camera movement with all three parts: motion type, amplitude, speed.
  ("slow push-in, small amplitude" — not "camera moves".)
- Write dialogue as: (S1) speaks softly, [English] the line here.
- Every detail must correspond to something visible or audible. No plot
  summaries. No unresolved reference tags.
- All timing must fit within the requested duration.
- Stay under 7,000 characters.

Target duration: {duration} seconds. Aspect ratio: {ratio}.

What the handoff actually costs

Almost nothing, which is the part that makes it practical. A typical handoff sends about 1,500 tokens in and gets about 2,500 back — roughly $0.0035 at M3's list rates. The five-second 768P clip it describes costs $0.40. Writing the prompt costs about 1% of generating the clip, so there is no reason to hand H3 a rough sentence to save money. Estimated from official list prices on 14 August 2026.

The prompt generator is the productised version of this, and reference to video is where images, video and voice go in.

The questions behind the mix-up

Is MiniMax M3 a video model? No. It reads video and describes it, but outputs text. For video, use H3.

Is M3 newer or better than H3? Neither. Different lines, two months apart. The numbers are not comparable.

How many parameters does M3 have? About 428B total, 23B active, per the model card. Hugging Face metadata says 427,040,140,160.

Is Hailuo 3.0 the same as H3? Yes. All three spellings mean MiniMax H3. Hailuo AI is the app, not the model.

Can I use M3 to write H3 prompts? Yes. MiniMax's own guidance invites developers to build their own prompt preprocessing.

Are both open weights? Both, under restricted licences. H3 excludes four territories; M3 excludes none but is non-commercial by default.

Which one should I pay for? Whichever matches your output. Video bills per second, text bills per token.

Why does M3 show up when I search for video? Its model card is tagged video — for understanding video. That plus one letter is enough.

Is Kling O3 a MiniMax model? No. Kuaishou makes it and calls it Video 3.0 Omni. It only looks similar.

I just want to make a video. Where do I start? Here — no M3 needed. For all the routes in, see how to use MiniMax H3.

What may change. MiniMax ships fast: check the release notes for anything newer than M3 and H3. M3's current rates are labelled a permanent 50% discount, so the cost comparison moves if that ends. And BrowseComp 83.5 is third-party reported; we have not found it in MiniMax's own text.