Which MiniMax H3 model should you use?

Posted November 3, 2026 by
Try MiniMax H3 on Replicate

Run H3 Highspeed

MiniMax’s H3 family has three video models on Replicate: minimax/h3-highspeed, minimax/h3-lite, and minimax/h3. A 10-second clip at 768P costs $0.60, $0.40, or $0.80 depending on which one you pick. Our suggestion is to start with Highspeed, drop to Lite when you’re generating lots of options, and move up to H3 when you need 2K or the shot is hard.

ModelResolutionsPrice per output second10 seconds at 768PGood for
H3 Highspeed768P$0.06$0.60Everyday work and reference-driven iteration
H3 Lite480P, 768P$0.025 at 480P, $0.04 at 768P$0.40Exploring ideas and larger batches
H3768P, 2K$0.08 at 768P, $0.13 at 2K$0.80Final shots and 2K delivery

Prices are from the model pages on October 31, 2026. At 2K, a 10-second H3 clip is $1.30.

The three share the same inputs otherwise. They all make clips from 4 to 15 seconds, and they all do text-to-video, image-to-video from a first and optional last frame, and generation guided by image, video, or audio references.

The example videos below come from the model pages. They use different prompts, and one of them starts from an image, so you can’t line them up as a fair comparison.

Start with Highspeed

Highspeed is the one to iterate with. MiniMax pitches it as a balance between quality and speed, and it keeps the full set of reference inputs. It only outputs 768P.

A hummingbird hovers beside a bright orange hibiscus flower in a sunlit garden, wings a soft blur, shallow depth of field, slow push in

Text-to-video, 768P, 5 seconds. The run took 21 seconds. (See the run.)

References are where Highspeed earns its place. Say you’re making a short fashion spot and you have one image for the location, one for the character, and one for the bag. Tell the model what each image is for: which one sets the scene, which one is the character, and what has to stay the same from shot to shot. Then describe the action and the camera. There’s a full prompt for this further down.

Use Lite when you want more attempts

Lite is the cheapest of the three, at $0.04 per second at 768P and $0.025 at 480P. Use it for storyboards, alternate openings, and any batch where you plan to throw most of the results away.

A hundred 10-second clips at 768P come to $40 on Lite, $60 on Highspeed, and $80 on H3.

Pick Lite to save money, not time. On the model-page examples, a 5-second Lite clip and a 5-second Highspeed clip both took about 21 seconds, though one run each proves very little and run time depends on the prompt and on load. Lite is also the model where you should look hardest at whether it did what you asked. A clip can look great and still have swapped the wrong object or dropped a character. If your batch involves fiddly edits, try a few on Highspeed first and see whether the difference matters to you.

Choose H3 for demanding final shots

H3 is the top of the family, and the only one of the three that outputs 2K. It’s also the only one whose resolution defaults to 2K, which matters if you’re watching costs.

the car revs through the desert

Image-to-video from a square first frame, 2K, 5 seconds. The output is 1440 × 1440 because it keeps the input image’s shape. The run took about three minutes. (See the run.)

Reach for H3 on motion graphics, choreographed effects, and edits where several things have to be right at the same time. Picture a logo animation where a dot turns into a letter, then an orbit, then the final mark. If one step is out of order or the text comes out misspelled, the clip is unusable. 2K won’t fix lettering for you, so check the text on every take.

Animating stills with H3

H3’s first_frame_image input turns a still into a moving shot, and it’s where the model’s 2K output earns its price. We took seven stills, a mix of photos and illustrations, and gave each one to H3 with a short note about what should move. Every clip below is 5 seconds at 2K, which costs $0.65.

The output keeps the shape of the input image, so a 4:3 still comes back at 1888 × 1440 and a 16:9 one at 2560 × 1440. The prompts mostly describe motion and camera. The image already covers what’s in the shot.

Bedroom

Slow, steady dolly push-in toward the bed. Soft daylight shifts gently across the wall as a cloud passes. The rubber plant’s leaves sway slightly in a light breeze from an open window, and the green throw ripples a little at its edge. Calm, quiet, no people.

Took 2 minutes 47 seconds. (See the run.)

Tiger

The tiger slowly prowls forward through the undergrowth toward the camera, shoulders rolling with each step, then opens its eyes and looks straight into the lens. Ferns and grass blades brush against its fur and spring back. Keep the painterly illustrated style with visible brush strokes. Camera slowly pulls back to keep pace.

Took 2 minutes 37 seconds. (See the run.)

Road trip

The car cruises down the highway at golden hour. Wind streams through the open window and whips the woman’s long auburn hair around her face. Her eyes stay open and bright the whole time; she smiles, laughs softly and keeps looking at the driver, who glances back at her. Sunlight flickers through passing trees, flaring across the windshield. Handheld camera from the back seat with gentle road vibration. Warm 1970s film look.

Took 4 minutes 23 seconds. (See the run.)

Crosswalk

Static high-angle camera. The cyclist actively pedals the red bicycle across the crosswalk: his legs push the pedals around in clear, continuous circular strokes, knees rising and falling, the cranks and wheels visibly turning as the bike rolls forward through the shaft of golden sunset light. His long shadow on the stripes shows the legs pumping too. The pedestrian in yellow keeps walking up and out of the light. Strong contrast, cinematic street photography look.

Took 2 minutes 48 seconds. (See the run.)

Subway

The subway car rocks gently as it moves, fluorescent lights flickering, passing tunnel lights sweeping across the windows. The woman holds her glance over her shoulder, then turns to the man and smiles; he looks down at her and grins. Passengers in the background sway. Handheld 35mm film camera feel, soft grain.

Took 3 minutes 20 seconds. (See the run.)

Diner

Rain streams down the diner window, droplets running and merging. Blurry headlights pass outside. The neon sign flickers once. The woman slowly lifts her coffee mug, takes a sip, and sets it down, gazing out at the rain. Steam curls from the mug. Very slow push-in. Moody teal and red film look.

Took 3 minutes 7 seconds. (See the run.)

Horses

The four horses gallop left to right across the meadow at full speed, manes and tails flying, hooves kicking up dust. The camera pans smoothly with them so the background streaks into horizontal motion blur. Black and white, grainy monochrome film.

Took 3 minutes 20 seconds. (See the run.)

The painterly tiger stays painterly, and the black and white horses stay black and white without any extra prompting. All seven took between two and a half and four and a half minutes, about what you’d expect from the desert car example above.

Three prompts to try

A fashion spot with four references

Use Highspeed. Pass four reference images in this order: the setting and mood, the character, the bag, the closing logo.

Create a 15-second, 16:9 fashion campaign. Use reference image 1 for the desert-road setting and restrained film texture, image 2 for the character, image 3 for the black bag, and image 4 for the closing logo. Beside a vintage car, the woman returns to the trunk, opens it, takes out the bag, briefly exchanges a look with the man by the car, and walks away. Keep the bag’s shape and the characters’ appearance consistent. Use purposeful cuts and a cool, understated mood. End on the supplied logo.

The bag is the thing to watch. If its shape drifts between cuts, the spot doesn’t work.

A seaside market that turns into a sketch

Use Lite to try a few versions, then run the one you like best on Highspeed.

Create a 15-second, 16:9 handheld-phone video of a bright seaside market, using four shots joined by brisk hard cuts. Include pale stalls, awnings, shell trinkets, an iced-drink stand, curtains moving in the breeze, and a blue sea in the distance. A right hand holding a thin cyan pen enters from the bottom of the frame. In each shot it points to one object. Only that object transforms into rough hand-drawn 2D animation; the rest of the scene stays live action.

Only the object the pen points at should turn into a drawing. When a version gets that right, keep the prompt and change one thing at a time.

A brand animation

Use H3, at 2K if that’s what you’re delivering.

Create a 15-second flat motion-graphics film for MiniMax H3. The only readable text is “MiniMax H3” and “Intelligence With Everyone”. Build the animation from a small solid dot, an H-shaped bridge, and three orbital tracks. Each shape must evolve continuously into the next, with a clear visual connection between transformations. Keep the visual system consistent and finish on a stable, legible brand frame.

Writing “2K” in the prompt does nothing. Set the resolution input.

Run Highspeed with the API

Install Replicate’s Python client, set REPLICATE_API_TOKEN, and run:

import replicate

output = replicate.run(
    "minimax/h3-highspeed",
    input={
        "prompt": (
            "A 10-second product film of an unbranded black travel bag "
            "on a pale stone pedestal. Begin with a close-up of the "
            "fabric texture, then slowly pull back to reveal the whole "
            "bag. Soft side lighting, restrained camera movement, "
            "consistent bag shape, no text or logos."
        ),
        "duration": 10,
        "ratio": "16:9",
    },
)

print(output)

For Lite, change the model to minimax/h3-lite and add "resolution": "768P" or "480P". For H3, use minimax/h3 with "resolution": "768P" or "2K". Set it explicitly when you’re comparing, since H3 defaults to 2K and Lite defaults to 768P.

To start from an image, pass first_frame_image, and last_frame_image if you want to pin the ending. For references, use reference_image_urls (up to 9), reference_video_urls (up to 3), and reference_audio_urls (up to 3). MiniMax requires an image or video reference alongside any reference audio. Each model’s API page lists the full set of inputs.

Comparing them yourself

Run the same prompt, references, duration, and aspect ratio at 768P on all three, a few times each. Then score what your project cares about: whether the subject stays recognizable, whether the camera does what you asked, whether text comes out right. Time the runs separately from the length of the clips.

If you’re delivering at 2K, test H3 at 2K on its own terms. A 2K clip next to a 768P clip tells you which one you’d ship, not which model is better.