MiniMax H3 Max vs MiniMax H3

The Short Answer

If you are delivering at 768p or below and want many takes, use H3 Max. If you need 2K, reference inputs, or the ability to run the model yourself, use MiniMax H3. That is the whole decision.

MiniMax H3 Max is fal's post-trained version of MiniMax H3. It renders a 5-second clip at 768p in under 3 seconds, costs $0.06 per second at that resolution, and comes back with synchronized audio in the same pass. It is hosted only — the weights are not released.

MiniMax H3 is the base model MiniMax open-sourced. It reaches 2K, adds reference-to-video and video editing, and you can download the weights and run it on your own hardware. It is substantially slower on the same job.

Neither is a setting of the other. They are separate models with separate endpoints, which is the single most common misunderstanding about the pair — including in the MiniMax H3 Max AI video generator most people land on first.

MiniMax H3 Max vs MiniMax H3: Specs Side by Side

Every axis where the two differ, and the three where they are identical. Figures are fal's published specs, checked 2026-08-28.

Three rows are worth reading twice. Reference to video and video editing exist only on the base model — if your workflow depends on either, the decision is already made. Open weights is the other hard split: base H3 can be downloaded and run locally, H3 Max cannot, and no amount of budget changes that.

The three identical rows matter too. Both models run 5 to 15 seconds, both are locked to 24 FPS, and both return synchronized audio in the same pass rather than as a second step. So if someone tells you the base model has better sound, or that H3 Max is limited to shorter clips, neither is true — those axes are the same on both.

Full resolution and aspect-ratio detail lives on the H3 Max resolution page.

Test 1 · 16:9

Pit stop — timed action in one continuous shot

What to inspect: Sequence order and simultaneous action

View the shared prompt+

Prompt summary · applied to both clips

Create a five-second 16:9 broadcast-style motorsport shot from a fixed low pit-wall camera. First, an orange-and-gunmetal prototype brakes into its box. Next, exactly four crew members replace all four wheels and release the front jack. Finally, the car launches left through tyre smoke. Use hard midday light, heat shimmer, synchronized wheel-gun, engine, nut, and radio audio. Keep every team and sponsor fictional; no collision, fire, or extra crew.

MiniMax H3768p · 5s

H3 Max768p · 5s

Watch whether arrival, wheel change, and launch remain three distinct beats — and whether four crew members stay coherent while working at once.

Test 2 · 16:9

Motion poster — typography under a timed hold

What to inspect: Spelling, layout, and final-frame stability

View the shared prompt+

Prompt summary · applied to both clips

Create a five-second 16:9 science-fiction teaser. Rise slowly past a ruined megastructure through ash, volumetric light, restrained cyan accents, and subtle camera roll. At three seconds, cut to black. During the final second, hold exactly two centered lines of white geometric type: “THE LONG QUIET” and “IN THEATRES 07 . 24”. Add a sub-bass rise, a distant structural groan, and one impact as the title appears. Include no other text, logo, watermark, subtitle, dialogue, or continuous score.

MiniMax H3768p · 5s

H3 Max768p · 5s

Look at the last second: the title, release line, letter spacing, and stillness make text errors easy to compare side by side.

Test 3 · 16:9

Creature reveal — crane-to-push camera move

What to inspect: Camera continuity, scale, and subject stability

View the shared prompt+

Prompt summary · applied to both clips

Create a five-second 16:9 photoreal cinematic shot in one take. Begin high above a research trawler in a heavy dusk swell, then crane toward the waterline as a vast dark, ridged creature rises behind it. Transition into a low push across the water while the ship becomes tiny against the silhouette. Keep the creature mostly unlit and clearly separated from the vessel. Use storm wind, hull impacts, cascading water, one deep pulse, and a distant interrupted shout; no music.

MiniMax H3768p · 5s

H3 Max768p · 5s

Compare the handoff from the descending crane move into the low push, then check whether the vessel and rising form keep a believable scale relationship.

Test 4 · 9:16

Creator footage — dialogue, autofocus, and drone launch

What to inspect: Lip-sync, phone-camera behavior, and vertical framing

View the shared prompt+

Prompt summary · applied to both clips

Create a five-second vertical 9:16 phone video on a windy beach at golden hour. A young man records himself, delivers one short frustrated line about a third attempt, then turns the phone toward an unbranded compact drone on wet sand. Its arms open, the rotors accelerate, and it exits upward while he follows it and laughs. Include one brief autofocus hunt, a natural exposure shift, handheld wobble, wind-distorted speech, surf, and rotor sound. No music or burned-in captions.

MiniMax H3768p · 5s

H3 Max768p · 5s

Listen to the spoken line against wind noise, then watch for one deliberate focus hunt and a clean transition from selfie framing to the drone launch.

Test 5 · 16:9

Papercraft spot — keeping a physical style intact

What to inspect: Material consistency and readable generated type

View the shared prompt+

Prompt summary · applied to both clips

Create a five-second 16:9 tabletop advertising shot made entirely from layered cut paper with visible fibres, gouache color, real edges, warm practical light, and stepped motion. A coral-and-cream paper aircraft crosses layered clouds; the clouds reveal a paper coastline, lighthouse, waves, and wired gulls; then a lower-third card rises with the words “FLY SOUTH”. Use paper rustle, card friction, soft felt impacts, brushed-brass music, and a final glockenspiel note. Never drift into photography, vector art, cel animation, or 3D rendering.

MiniMax H3768p · 5s

H3 Max768p · 5s

Follow the paper fibres, cut edges, layered shadows, and closing banner. The useful difference is whether the handmade look survives once everything starts moving.

Test 6 · 16:9

Dive watch — macro geometry and sound cues

What to inspect: Audio-to-action timing across three precise cuts

View the shared prompt+

Prompt summary · applied to both clips

Create a five-second 16:9 black-on-black macro product sequence with a 100 mm lens and one raking key light. First, gloved fingertips pull an unbranded crown through two visible clicks. Next, a steel bracelet clasp closes on slate with one sharp snap. Finally, the dark dial fills the frame while the sweep hand makes eight discrete steps across the six-o’clock index as the light fades. Align every click, snap, and tick to visible contact. Use only dry studio sound; no music, voice, dial text, or extra hands.

MiniMax H3768p · 5s

H3 Max768p · 5s

Match the two crown clicks, clasp snap, and eight dial ticks to the visible contact points while checking whether the watch geometry holds in macro view.

Source clips and test conditions: fal's MiniMax H3 vs. MiniMax H3 Max comparison. Videos are served from local copies so the comparison does not depend on third-party playback availability.

Where MiniMax H3 Max Wins

Three advantages, and only one of them is the number you would expect. Speed matters less as a spec than as a change in how you work.

Speed, measured against the whole field

Design Arena placed H3 Max on the Pareto frontier for video generation: no other image-to-video model on that arena delivers higher preference at a lower generation time. Its measured runs came in at 6.4 seconds for image to video, 18× faster than the arena average, and 4.7 seconds for text to video, 24× faster. Against the base model specifically, Design Arena measured roughly 50× faster.

Price at 768p

H3 Max lists at $0.06 per second at 768p — the lowest listed API price of any model at the top of the Artificial Analysis image-to-video board. On a 15-second clip that is $0.90 against $1.50 for Wan 3.0 and $7.10 for Seedance 2.5 at their nearest tiers.

Prompt adherence and aesthetics

Those two are what fal's post-training targeted first, and the boards reflect it: Design Arena scores H3 Max at 1,341 Elo against the base model's 1,333. That is an 8-point gap — real, but narrow. Treat it as "slightly ahead", not "in a different class".

Where MiniMax H3 Wins

This site runs H3 Max, so read this section knowing that. There are three jobs where the base model is simply the right call, and pretending otherwise would waste your time.

You need 2K

H3 Max stops at 768p. It is a hard ceiling, not a setting. If your deliverable specifies 2K, the base model is the only one of the two that gets there, and no workaround on our side changes that.

You need reference inputs or editing

Base MiniMax H3 takes reference images, video, and audio into a single context, and it can edit existing clips. H3 Max shipped with text-to-video and image-to-video only; reference-to-video was announced as coming later. For character sheets, style references, or re-cutting footage you already have, that is the base model's job.

You want to run it yourself

The MiniMax H3 open source release is the whole reason this axis exists: you can download the base weights and run them on your own hardware, and people do, in ComfyUI and elsewhere. H3 Max is hosted only and its weights are not published. No pricing tier or support plan changes that, because the files simply are not out there. If self-hosting is a requirement rather than a preference — an air-gapped environment, a compliance rule, or just a preference for owning the stack — this page ends here and the base model is your answer.

What Each One Actually Costs You

Every other comparison of these two reports API rates per second. Here is the same thing in clips and dollars, which is the unit you actually buy in.

On this site H3 Max is billed in credits: 480p costs 1 credit per second, 768p costs 2. So a 5-second clip at 768p is 10 credits, and the longest single generation — 15 seconds — is 30.

Sign-up credits cover one 5-second clip at 768p, which is enough to judge the audio and the prompt adherence before you decide anything. Full pack detail is on the MiniMax H3 Max pricing page.

The base model is priced separately and per its own tiers, so a like-for-like number depends on which provider you run it through. The honest summary: at 768p H3 Max is the cheaper of the two, and the gap widens the more takes you run.

Which One to Pick

Four situations, four answers. Find the one closest to your work and stop reading — the rest of this page is detail you do not need.

Social, ads, and previz at 768p → H3 Max

Short clips, many takes, delivery on a feed or a deck. Nothing here needs 2K, and the three-second turnaround changes how many versions you try. This is the case H3 Max was built for.

Broadcast or client delivery specified at 2K → MiniMax H3

If the spec sheet says 2K, the decision is made for you. Budget the extra render time.

Character sheets, style references, or re-cutting existing footage → MiniMax H3

Reference-to-video and editing are base-model-only. Check back on H3 Max later — reference-to-video was announced as coming — but today the base model is the one that does it.

Self-hosting, air-gapped, or you just want the weights → MiniMax H3

Only the base model is downloadable. This is the cleanest split of the four and it has nothing to do with quality.

If you landed in the first bucket, you can start generating with H3 Max on sign-up credits.

MiniMax H3 Max vs MiniMax H3 FAQ

Short answers first. Every figure is fal's published spec or a linked third-party leaderboard, checked 2026-08-28.