What Is H3 Max? A Faster MiniMax H3, Explained

H3 Max is a video model that renders a clip faster than you can watch it. A five-second shot at 768P comes back in under three seconds, with sound already in it.

The one-line difference from every video model before it: you stop submitting jobs and start steering. There's no queue to check and no tab to leave open — you type, look, and change it.

A note on the numbers below: the timings are the published figures and arithmetic off them, not head-to-head runs of my own. Where something is a claim rather than a measurement, it says so.

First, the name

H3 Max is not a MiniMax release. It is a post-trained version of MiniMax H3, built from H3's open weights with the MiniMax team involved, and shipped on 1 September 2026.

Four names are now circling the same family, and search results mix them freely:

• MiniMax H3 (also Hailuo 3) — the original model from MiniMax, 4–15 seconds, up to 2K.
• H3 Max — the post-trained variant. Faster, capped at 768P. What this article is about.
• H3 Max Live — the branding used for the launch-week demo where a stream generated continuously.
• H3 Max Director — the continuous model behind that demo. It shipped as its own endpoint days after H3 Max, and it is not "H3 Max but longer."

All of them are callable now. The split that matters is clips versus stream: H3 Max returns a finished file, Director holds an open session that you steer while it runs. They bill on completely different meters, which the cost section gets into.

The open weights are why this exists at all. MiniMax published H3's weights; the Max variant post-trained them and rebuilt the inference path around the result. The post-trained result now sits a few Elo above the model it came from on Design Arena; the weights it started from are the H3 open weights release.

What you can actually make with it

The output is a 5–15 second clip at 480P or 768P with synchronized audio — the same family of output as H3, one notch shorter at the bottom and one tier lower at the top. What changes is what you can do with a shot that arrives in three seconds instead of the minute and a half or two the official endpoint takes.

Iterated single shots. This is the honest headline use. A shot that needs eleven attempts to get the camera move right used to be twenty minutes of submitting and tab-switching; now it's a few minutes of sitting there adjusting one clause at a time. It helps most when you can't specify the shot in advance.

Directed continuous video. This is the Director endpoint, and it is the genuinely new thing. Rather than chaining discrete clips, it holds one video stream open and keeps context across it, so you can send fresh prompts mid-stream and the character, setting and continuity survive the change. Sessions run up to two minutes as standard, with longer ones gated behind a request. You can seed it with a first frame. This is what the early-September live-stream demos were running on, and the cost section below is the part those demos leave out.

Ecommerce variant sets. Fifteen versions of one product shot — different angles, different lengths, different placements — stop being an overnight batch. At three seconds each you generate the set, look at all fifteen, and rerun the four that missed.

First-to-last keyframe shots. The image-to-video endpoint takes both image_url and end_image_url, so you can pin where a shot starts and where it ends. This is the same technique covered in our first and last frame guide and its prompt builder, and it is the most reliable way to get a specific movement instead of a plausible one.

Rough cuts for longer pieces. Nothing here raises the 15-second ceiling. A 90-second piece is still six or more generations stitched by you. What the speed changes is that you can block out all six in one sitting and find out your sequence doesn't work before you have paid for 2K renders of it.

The ceiling it does not lift: 768P. H3 Max has no 2K mode.

What "faster than real time" actually costs

The published list price is $0.08 per second of video at 768P, or $0.05 per second at 480P — quoted elsewhere as $4.80 per minute at 768P. Introductory discounts were running at publication and expire on different dates — 7 September for the clip endpoints, 14 September for Director — so re-check the rate before you budget anything.

At list price, per clip:

That is cheap for a shot. Now hold it against the thing everyone wants to build.

The Director meter has a floor the clip meter doesn't

Director lists at the same $0.08 per second of generated video. The rate is not the difference — these two clauses are:

• Every session bills a minimum of 60 seconds, whatever the viewer actually watched. Someone who opens your stream, looks at four seconds and closes the tab still costs a full minute.
• Sessions cap at two minutes as standard. Longer runs need approval for a named use case.

So the smallest possible unit of Director is a minute of video — $4.80 at list price, and there is no way to sample it more cheaply than that. Any "try it free" button wired to Director is a $4.80 button.

A continuous stream bills per second of output, not per second of compute. If a channel plays generated video around the clock, it consumes 60 seconds of generated video for every 60 seconds on the wall. The run rate:

The stream demos going around are sponsored by the company that sells the inference. That is not a criticism — they are proving a capability, and the capability is real. But if you are reading them as a business you could clone this weekend, the row above is what stands between you and that. A free public channel with no per-viewer revenue burns about $48,000 a week.

The version that survives contact with a bill is a metered session: one viewer, their own stream, their own credits ticking down, and a minimum charge that matches the 60-second floor you are being billed on. Same capability, meter attached to the person watching. Worth settling before you write any code, because the 60-second minimum makes the usual free-trial shape unaffordable and you will want a different hook.

How it actually runs

Two things were changed, not one. The open weights were post-trained for prompt adherence and aesthetics, and the serving path was rebuilt around that specific model. The published figure is 35x the throughput of the official H3 endpoint, up to 50x at best. Those numbers come from the party that stands to gain from them, measured on their own benchmark; treat them as a claim, not a measurement anyone here made.

prompt_expansion_mode is the setting that will bite you. It takes balanced or quality. The published schema puts balanced at about a second of overhead and quality at about thirty. Thirty seconds is ten times the render itself — set it to quality by default and you have thrown away the entire reason you came here. Steer on balanced, switch to quality for the take you intend to keep.

Aspect ratio works differently on the two endpoints. Text-to-video takes an aspect_ratio parameter covering 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Image-to-video has no such parameter — it inherits the ratio of the image you pass in. If you want vertical output from an image, the first frame has to already be vertical. This catches people who assume the parameter is just hiding somewhere.

Sound arrives with the picture. Every generation returns synchronized audio — room tone, foley, ambience, music — in one pass rather than as a second job. The 16:9 setting at 768P comes out 1344×768 at 24fps.

The benchmark placings are real but narrow. Checked 5 September 2026, H3 Max sits first on Design Arena's image-to-video board at 1,341 Elo, with MiniMax H3 right behind at 1,333, and first on Artificial Analysis's image-to-video-with-audio board at 1,201 Elo across 2,177 samples. Eight Elo points over its own parent is a lead, not a gulf. The speed is the story; the quality is "it did not regress."

Where it falls short

No 2K. 480P and 768P, and that's the list. It's the main reason to keep plain H3 in the pipeline rather than replacing it.

Five seconds is the floor, fifteen the ceiling. The duration default is 5. Short beats — a two-second reaction, a one-second insert — mean generating five seconds and trimming.

Reference inputs arrived late and are still thinner than H3's. H3 accepts up to 12 files in one request: 9 images + 3 videos + 3 audio. H3 Max launched with text-to-video and image-to-video only; reference-to-video followed within days. If your work leans on feeding a reference video and matching its performance, check what the endpoint accepts today rather than assuming parity with the original.

The prompt length limit is not published. H3 documents roughly 7,000 characters. The H3 Max schema does not state a limit, so a very long brief is something to test rather than assume.

Introductory pricing expires. Anything you cost out this week at the promotional rate needs re-checking against $0.05 and $0.08 before it becomes a plan.

Director's two-minute cap is the default you will actually be on. Longer sessions exist but need a named use case approved, so "infinite" in the demos means sessions restarting rather than one unbroken run. Past two minutes you are back to stitching — same seam problem as chaining clips, just less often.

When you still want plain H3

Reach for the original when the deliverable is 2K, when you need the full 12-file reference set, or when the shot is going to be right on the first or second try anyway — in that case the speed advantage buys you nothing and the resolution loss costs you something real.

The workflow that gets the most out of both: block and iterate at 480P, final-render on H3 at 2K. 480P is the only tier where H3 Max is actually cheaper rather than just faster. Find the shot there, then re-run the winning prompt on H3 for delivery — the prompts transfer, because it's the same weights underneath.

H3 runs in the browser here with free credits on signup, no install — and our H3 Max page has the sample reel and the settings that differ. If the piece is going to run longer than 15 seconds, the free H3 long video planner does the segment math first, which is cheaper than discovering the problem after six generations.

Things people mix up

H3 Max is not an upgrade to H3. It's a sibling with a different trade: faster across every tier, cheaper only at 480P, and no 2K. "Max" reads like the bigger model in a lineup, and here it isn't.

"Faster than real time" is about generation, not playback. A five-second clip rendering in three seconds means one machine can produce video faster than one screen can play it. It does not mean the video streams as it generates in your own app for free.

"Director" and "H3 Max" are not the same purchase. Director is a separate endpoint with a session meter, a 60-second minimum and a two-minute cap. Code written against the clip endpoints does not become a live stream by swapping a model name.

H3 Max is not a MiniMax release. MiniMax made and open-weighted the original; the Max variant was post-trained from those weights by someone else. Credits, accounts and pricing on the two are unrelated.

Is it worth it

If you iterate — if your process is generate, look, adjust, generate — H3 Max changes the shape of your afternoon more than any quality bump this year did. Three seconds is under the threshold where you go do something else, and staying in the loop is what makes prompts get better.

Deliver at 2K, or lean on reference videos and audio, and it is a drafting tool rather than a replacement — plain H3 stays in the pipeline.

The live-stream demos are a capability proof, not a business plan. Do the run-rate arithmetic before you design around them.

For the model underneath, see what MiniMax H3 is and the H3 API guide. For a spec-by-spec look at what you give up moving across, we put the two side by side in H3 Max vs MiniMax H3. Either way the H3 prompt guide sets the ceiling — a faster loop only helps if the prompts improve.

This site is an independent third-party tool, not affiliated with MiniMax; credits here don't transfer.