MiniMax H3 MODEL GUIDE

MiniMax H3: Features, Examples, Prompts, and Specs

A practical guide to MiniMax H3 inputs, native sound, multimodal controls, output options, open weights, and the workflows available in Minimax Motion.

Updated August 15, 2026

Try MiniMax H3 free

MiniMax H3

What is MiniMax H3?

MiniMax H3 is a general-purpose multimodal video model from MiniMax. It can interpret text, images, video, and audio in one context and generate video with native stereo sound.

MiniMax H3 specifications

Choose any whole-second duration from 4 to 15 seconds.

MiniMax H3
4–15s
Output options
768P · 2K
IMAGES · VIDEO · AUDIO
9 · 3 · 3
Native stereo sound
Stereo

Choose the input that fits the shot

Core MiniMax H3 capabilities

Unified multimodal context

Combine text, image, video, and audio references so subject identity, motion, camera language, timing, and sound can inform one generation.

Native stereo sound

Generate dialogue, voiceover, ambience, effects, and music in sync with the image instead of adding a separate audio pass.

Multi-shot storytelling

Describe timed shots and transitions in one prompt while keeping the overall subject and sound direction connected.

Reference-led control

Use references for appearance, composition, motion transfer, pacing, and audio reuse. Results still vary and should be reviewed before production use.

H3-Context-IR prompt interpretation

H3-Context-IR is a hosted asynchronous context-interpretation step. It returns a structured enhanced prompt for a later generation request; it does not create or edit video itself.

Up-to-2K regeneration

MiniMax offers a hosted 2K regeneration path. Availability can depend on the workflow and service used.

MiniMax H3

Eight MiniMax H3 workflow examples

Each example includes the inputs, prompt, settings, transcript when applicable, and a short note about the observed result.

Text·10s·768P·16:9

Rainy city tracking shot

A text-led street scene with a moving subject, camera tracking, reflections, and native rain ambience.

View prompt and settings

Inputs

  • text: Prompt

Prompt

integrated_multimodal_description: A continuous cinematic 10-second rainy-night city scene. A fictional clearly adult East Asian woman in a tailored charcoal coat walks calmly beneath a transparent umbrella along an original unbranded metropolitan street. 0-3s: medium-wide establishing view with wet asphalt and cool cyan and warm amber reflections; the camera begins a smooth parallel track. 3-7s: track beside her at walking pace while rain beads on the umbrella and her coat moves naturally; preserve face, hands, wardrobe, and street geometry. 7-10s: ease into a closer three-quarter view as she briefly looks toward the reflected lights, then holds the final composition. Natural motion, stable anatomy, restrained depth of field, no cuts, no readable signs, no logos. overall_soundscape: Native stereo rain, soft footsteps on wet pavement, subtle umbrella patter, and distant original city traffic moving across the stereo field. non_diegetic_music: A sparse original low ambient pulse, quiet beneath the rain, with no recognizable melody.

Settings

10s · 768P · 16:9

Transcript

[Rain falls as footsteps move along a wet street.]

Observed result

Focuses on coordinated subject and camera motion in a detailed night environment.

Text·10s·768P·16:9

Astronomer with native dialogue

A cinematic observatory scene combining a speaking character, timed dialogue, room tone, and camera movement.

View prompt and settings

Inputs

  • text: Prompt

Prompt

integrated_multimodal_description: A continuous cinematic 10-second scene inside an original mountaintop observatory at blue hour. One fictional clearly adult white woman astronomer (S1), wearing practical navy workwear, stands beside a large unbranded telescope. 0-3s: slow dolly from the instrument toward S1 as the dome slit reveals a dense natural star field. 3-7s: S1 turns from the eyepiece toward camera and speaks with restrained wonder, accurate lip sync and natural facial motion: <d>[English] The oldest light is still arriving.</d> 7-10s: she looks back toward the sky while the camera settles; preserve identity, telescope geometry, and realistic hands. No cuts, no readable interface text, no logos. overall_soundscape: Native stereo observatory room tone, a quiet servo movement, faint mountain wind outside, and the clearly recorded voice of S1 centered in the mix. non_diegetic_music: One original low sustained synth chord that blooms gently after the dialogue.

Settings

10s · 768P · 16:9

Transcript

[An observatory motor hums softly.] The oldest light is still arriving. [Mountain wind and a low musical tone.]

Observed result

Focuses on speech timing, facial motion, and a soundscape generated with the picture.

Frames·10s·768P·16:9

Fashion image to motion

A first-frame workflow that carries the source styling into a directed fashion movement.

View prompt and settings

Inputs

  • image: First frame

Prompt

Use the supplied first frame as the exact opening composition and identity anchor. integrated_multimodal_description: A continuous 10-second editorial fashion shot. The fictional clearly adult East Asian woman in the original emerald sculptural coat walks through the pale stone arcade. 0-3s: begin pixel-aligned to the supplied frame; fabric and hair first respond to a light breeze. 3-8s: the camera performs a slow controlled clockwise orbit while she takes two natural steps and turns her shoulders toward the new camera angle. 8-10s: settle into a stable full-body profile with the same face, emerald coat, black clothing, footwear, architecture, sunlight direction, and proportions. Natural hands and feet, continuous motion, no cuts, no logos, no new people. overall_soundscape: Stereo footsteps on stone, soft fabric movement, and a light open-air breeze. non_diegetic_music: Minimal original brushed percussion with two soft accents.

Settings

10s · 768P · 16:9

Transcript

[Footsteps on stone, moving fabric, and a light breeze.]

Observed result

Focuses on preserving the source composition while introducing subject and fabric motion.

Frames·10s·768P·16:9

First-to-last-frame perfume transition

A product transformation guided by a defined opening frame and destination frame.

View prompt and settings

Inputs

  • image: First frame
  • image: Last frame

Prompt

Use the supplied first frame as the exact opening and the supplied last frame as the exact ending. integrated_multimodal_description: A seamless 10-second luxury product transformation with one fixed unbranded faceted perfume bottle. 0-2s: hold the clean start composition while a narrow amber rim light travels across the unchanged glass. 2-7s: the three dry ivory petals lift and multiply into fresh translucent amber and ivory petals; droplets rise in smooth arcs and a luminous amber ribbon grows behind the bottle. Keep the bottle, blank label, cap, plinth, camera, reflections, and background stable. 7-10s: all motion resolves precisely into the supplied last frame without a cut or crossfade. No text, no logos, no extra bottles. overall_soundscape: Close stereo glass resonance, soft petal movement, delicate droplets, and a warm rising air texture. non_diegetic_music: An original restrained three-note glass motif resolving on the final frame.

Settings

10s · 768P · 16:9

Transcript

[Petals stir as glass tones rise and resolve.]

Observed result

Focuses on the continuity of the transition and product readability between endpoints.

References·10s·768P·16:9

Fashion identity and motion reference

Separate visual and motion references direct the subject styling and performance of one shot.

View prompt and settings

Inputs

  • image: Picture 1
  • video: Video 1

Prompt

subject_definitions: Subject 1 is the fictional adult woman from Picture 1. Retain her invented face, short copper bob, freckles, cobalt sculptural jacket, charcoal trousers, body proportions, and shoes at high strength. Video 1 supplies walking cadence, body timing, and smooth parallel camera motion only; do not retain its person, wardrobe, umbrella, street, weather, or audio. summary: Create a 10-second original studio fashion shot in which Subject 1 performs the motion rhythm from Video 1 inside Picture 1's concrete studio. retention_analysis: Picture 1: retain Subject 1 identity and wardrobe strongly; retain the concrete studio, curtain, and window-light direction. Video 1: retain motion timing and camera rhythm only. Reject all other Video 1 appearance and sound. detailed_description: 0-3s: begin near Picture 1 with Subject 1 in a neutral stance. 3-8s: she performs the referenced step cadence and a controlled half-turn while the camera tracks in parallel, preserving face, hair, jacket structure, hands, and feet. 8-10s: she stops in a stable three-quarter pose as the camera eases to rest. One continuous shot, realistic cloth physics, no logos, no text, no extra people. overall_soundscape: Original stereo studio room tone, precise footsteps on concrete, and natural fabric movement; do not reuse Video 1 audio. non_diegetic_music: A quiet original metronomic pulse matching the motion without a recognizable melody.

Settings

10s · 768P · 16:9

Transcript

[Measured footsteps and fabric movement in a quiet studio.]

Observed result

Focuses on balancing appearance retention with movement borrowed from a reference clip.

References·10s·768P·16:9

Mixed image, video, and audio references

A multimodal scene directed with identity imagery, motion video, and an audio cue in one context.

View prompt and settings

Inputs

  • image: Picture 1
  • video: Video 1
  • audio: Audio 1

Prompt

subject_definitions: Subject 1 is the fictional adult East Asian sound designer from Picture 1. Subject 2 is the original suspended amber ribbon light sculpture in Picture 1. Retain both identities, the cream workwear, the dark acoustic studio, and the sculpture's material form strongly. Video 1 supplies only a smooth lateral-to-forward camera rhythm. Audio 1 supplies the timing, low-frequency pulse, and abstract tonal texture. summary: Create a 10-second original studio performance in which Subject 1 activates Subject 2 and the sculpture's light travels in exact sympathy with Audio 1. retention_analysis: Picture 1: retain Subject 1, Subject 2, wardrobe, studio, palette, and spatial relationship. Video 1: retain camera rhythm only; discard its identity, location, weather, props, and audio. Audio 1: strongly retain its pulse envelope and tonal texture as the timing source. detailed_description: 0-3s: begin close to Picture 1 as Subject 1 places one hand naturally on the generic control pad. 3-8s: the camera transitions from a gentle lateral move into a slow push while an amber light wave travels through Subject 2 in sync with Audio 1; Subject 1 follows it with her gaze. 8-10s: the wave returns softly and both camera and subject settle. Stable identity and anatomy, one continuous shot, no readable controls, no logos, no new people. overall_soundscape: Use Audio 1 as the audible foundation and add only subtle original switch contact and room reflections spatialized around it. non_diegetic_music: None beyond the supplied Audio 1 texture.

Settings

10s · 768P · 16:9

Transcript

[A low electronic pulse moves through the studio.]

Observed result

Focuses on how several labeled references contribute without replacing the main prompt.

Frames·10s·768P·16:9

Product and visible typography

A compact commercial shot designed around a hero object, controlled camera motion, and visible copy.

View prompt and settings

Inputs

  • image: First frame

Prompt

Use the supplied first frame as the exact opening composition and product anchor. integrated_multimodal_description: A continuous 10-second original beverage commercial. Preserve the same unbranded midnight-blue can and keep the generic visible word "MOTION" spelled exactly M-O-T-I-O-N, legible, unchanged, and appearing only once throughout. 0-3s: a cool edge light sweeps across the can while condensation gathers. 3-7s: the camera performs a slow 35-degree orbit; clear droplets arc behind the can and the pedestal rotates subtly while the label remains geometrically stable. 7-10s: return to a centered hero angle and hold the exact word cleanly. No other writing, symbols, logos, extra cans, or brand marks. overall_soundscape: Crisp original stereo droplets, a subtle aluminum resonance, and a low clean product impact at the final hold. non_diegetic_music: An original short electronic pulse with no recognizable melody.

Settings

10s · 768P · 16:9

Transcript

[Droplets, a metallic resonance, and an electronic pulse.]

Observed result

Focuses on product presentation and text stability; visible lettering should still be reviewed frame by frame.

Text·10s·768P·16:9

Multi-shot product and UI story

One timed prompt moves from a product reveal to interface detail and a closing campaign frame.

View prompt and settings

Inputs

  • image: First frame

Prompt

Use the supplied first frame as the exact opening product and interface design. integrated_multimodal_description: Create a coherent original 10-second, three-shot product/UI campaign using the same generic black smartphone, amber-and-violet ribbon motif, and unlabeled interface system in every shot. Shot 1, 0-3s: begin aligned to the supplied frame and dolly toward the device and two translucent cards. Match cut. Shot 2, 3-7s: macro close-up of the same practical timeline interface as one amber clip expands and the abstract ribbon preview responds; no readable words or nonsense text. Match cut. Shot 3, 7-10s: return to a clean wider hero composition as the physical amber ribbon light traces behind the same phone and the UI cards settle. Preserve device geometry, palette, interface hierarchy, and reflections across shots. No logos, trademarks, recognizable operating system, hands, or extra devices. overall_soundscape: Original stereo interface taps, gentle glassy transitions at both cuts, and a soft electrical ribbon sweep. non_diegetic_music: A concise original electronic bed with three structural accents aligned to the shot sequence.

Settings

10s · 768P · 16:9

Transcript

[Interface taps and glassy transitions over an electronic rhythm.]

Observed result

Focuses on shot transitions, visual continuity, and pacing across a short sequence.

Any whole second from 4s to 15s

Output options

Minimax Motion offers 4–15-second clips at 768p or 2K. Text generations support multiple aspect ratios; frame-led generations follow the source framing.

768P

Standard

2K

High detail

9 IMAGES · 3 VIDEOS · 3 AUDIO

Reference limits

Use up to 9 images, 3 videos, and 3 audio files. Reference videos and audio may total up to 15 seconds per media type. The creator validates formats, dimensions, duration, and file size before generation.

JPG / PNG / WebP / HEICMP4 / MOVWAV / MP3 / M4A

Limitations to plan for

Fine detail, readable text, identity, motion, dialogue, and reference adherence can vary between generations. Complex prompts may need iteration, and outputs should be checked before commercial delivery.

What is open-weight?

MiniMax has released the H3-Base weights under the H3 Community License. Review that license before local or commercial use.

What remains hosted?

Context-IR and the up-to-2K regeneration path are not included in the H3-Base weight release and remain hosted capabilities.

MiniMax H3 frequently asked questions

Is MiniMax H3 the same as Hailuo H3?+

MiniMax H3 is MiniMax's model name. Hailuo is a MiniMax video product name that may also surface MiniMax H3, while Minimax Motion is an independent interface for the model.

How much does MiniMax H3 cost here?+

Minimax Motion charges 6 credits per second at 768p and 12 credits per second at 2K. New accounts receive 30 free credits.

Can I run MiniMax H3 locally?+

H3-Base weights are available for local deployment subject to the H3 Community License and the required hardware. Hosted features are separate.

Does MiniMax H3 create native audio?+

Yes. MiniMax H3 can generate synchronized stereo dialogue, ambience, effects, and music as part of the video generation.

Does MiniMax H3 support 2K?+

MiniMax documents an up-to-2K regeneration path. Minimax Motion exposes a 2K choice for supported workflows.

How do MiniMax H3 references work?+

Label images, video, and audio consistently, state what each reference should retain, and describe their relationship to the requested shot.

MiniMax H3 is developed by MiniMax. Minimax Motion is an independent service and is not affiliated with or endorsed by MiniMax.

Try MiniMax H3 free
MiniMax H3: Features, Examples, Prompts, and Specs