70% off Wan3.0 Standard and 50% off Wan3.0 Prime. Try Wan3.0

MiniMax H3 is live on Runware

Access MiniMax H3, MiniMax's general-purpose video model that generates and edits clips from text, image, video, and audio inputs, with native stereo audio in every output. Five faster variants, from H3 Max to H3 Lite, run at 480p and 768p from $0.025 per second.

0:00

All workflows supported natively.

One integration for generation and editing, with no separate endpoints.

  • Text-to-video
  • Keyframe-guided video
  • Reference-to-video
  • Instruction-based editing

The MiniMax H3 lineup

MiniMax H3 is the full model, for finished work at up to 2K and for editing an existing clip. Five faster variants run at 480p and 768p for a lower price, each tuned for a different balance of speed, cost, and control.

ModelMiniMax H3H3 MaxH3 Max TurboH3 FastH3 HighspeedH3 Lite
Built forFinished work at up to 2K, and editing an existing clipFaster generation with strong prompt adherenceQuick iteration from text or framesFast previews and batches, with full reference controlEveryday and social content where turnaround mattersLow-cost previews, drafts, and batch variations
Resolution768p, 2K480p, 768p480p, 768p480p, 768p480p, 768p480p, 768p
Video duration4 to 15 seconds5 to 15 seconds5 to 15 seconds4 to 15 seconds4 to 15 seconds4 to 15 seconds
First and last framesYesYesYesYesYesYes
Reference inputs9 images, 3 videos, 3 audio clips9 images, 3 videos, 3 audio clipsNone9 images, 3 videos, 3 audio clips9 images, 3 videos, 3 audio clips9 images, 3 videos, 3 audio clips
Edits an existing clipYesNoNoNoNoNo
Native audioYesYesYesYesYesYes
Price from$0.08/sec at 768p$0.05/sec at 480p$0.025/sec at 480p$0.046/sec at 480p$0.04/sec at 480p$0.025/sec at 480p

MiniMax H3 pricing

Metered per second of output, by model and resolution. No subscription, no minimums, and the same key and billing as every other model on Runware, with no dedicated infrastructure to provision.

MiniMax H3 · Standard · 768p
$0.08/sec
MiniMax H3 · High resolution · 1440p (2K)
$0.13/sec

A 5-second clip costs $0.40 at 768p or $0.65 at 2K. Reference images: the first 5 are free, then $0.04 each. Reference video: $0.13 per second uploaded. Reference audio: free.

Variants, per secondH3 MaxH3 Max TurboH3 FastH3 HighspeedH3 Lite
480p$0.05/sec$0.025/sec$0.046/sec$0.04/sec$0.025/sec
768p$0.08/sec$0.04/sec$0.08/sec$0.06/sec$0.04/sec

Reference images cost $0.03 each beyond the first 5 on H3 Highspeed and $0.02 each beyond the first 5 on H3 Lite. H3 Fast, H3 Highspeed, and H3 Lite bill uploaded reference video per second, at their output rate.

Four ways to build a video

MiniMax H3 can generate an entirely new clip from a written prompt, follow frames pinned to specific points in a sequence, carry a subject in from reference inputs, or edit a clip you already have. All of it runs through the same Runware integration.

Text to video

A prompt alone is enough. MiniMax H3 generates an entirely new clip from scratch, with synced audio timed to the action as it happens.

Keyframe-guided video

Pin a first frame, a last frame, or both, and H3 fills in the motion that connects them. It reads closer to storyboarding than prompting.

Multi-reference conditioning

Combine reference images, video, and audio in a single call to carry a subject's look, motion, or voice into a brand new scene.

Instruction-based editing

Point at an existing clip, describe the change as an instruction, and get back the same shot with just that change applied.

Why MiniMax H3 stands out

Native audio, generated by default

Every clip returns with synchronized stereo audio grounded in the action on screen, timed to specific moments like a hammer strike or a footstep rather than layered on as generic ambience.

Keyframe-guided generation

Pin a first frame, a last frame, or both, and H3 generates the motion that connects them, closer to storyboarding than prompting.

Multi-reference conditioning

Combine reference images, video, and audio in one call to keep a subject's look, motion, or voice consistent as it moves into a new scene.

Instruction-based editing

Point at an existing clip, describe the change as an instruction, and get back the same shot with just that change applied, while the subject and camera path hold.

Six aspect ratios, at 768p or 2K

Renders at a fixed set of size pairs spanning cinematic 21:9 to vertical 9:16, at either resolution tier, so the frame is chosen before generation instead of cropped after.

Directed with cinematic language

Responds to camera, lens, and lighting terms like low-angle dolly-in or chiaroscuro lighting, so a shot can be directed rather than just described.

See it in action

Real generations and edits from MiniMax H3, each with what actually went into making it.

0:00
Edit

Day-to-night relight

Source clip
0:00
Text to video

Text-to-video with native audio

0:00
Native audio

Every hammer strike timed to the hit

0:00
Edit

Three edits, one pass

Source clip
0:00
Reference to video

A new subject, rendered in the reference's clay style

Style reference
Style reference

Built from references

Pass reference images, video, or frames alongside the prompt, and H3 builds the shot from them.

Motion transfer

The character performing the reference dance

The dance comes from one video and the dancer from one image. H3 carries over the performance, not the reference's own look.

0:00
Motion reference
Character reference
Character reference
0:00
Output
First and last frame

H3 generates the bloom between the two frames

Pin a first and a last frame, and H3 generates the motion that connects them: here, the bud opening into the bloom.

First frame
First frame
Last frame
Last frame
0:00
Output

How it compares

Where MiniMax H3 sits next to other leading video models on Runware. Its edge is native audio directed like picture, paired with a nine-image, three-video, three-audio reference surface and single-instruction editing.

ModelMiniMax H3Seedance 2.5Veo 3.1Kling VIDEO 3.0 Omni StandardGemini Omni Flash 1.2
Max duration15s30s8s15s15s (10s with first and last frames, references, or a source clip)
Input typesText, image, video, audioText, image, video, audioText, image, videoText, image, videoText, image, video
Reference inputs9 images + 3 video + 3 audio30 images + 10 video + 10 audioUp to 3 images7 images + 1 video + 7 elements10 images + 3 videos
Native audioYesYesYesYesYes
Max resolution1440p (2K)4K4K720p4K
CapabilitiesText-to-video, image-to-video, video-to-video, audio-to-video, editText-to-video, image-to-video, video-to-video, audio-to-video, edit, extendText-to-video, image-to-video, video-to-video, edit, extendText-to-video, image-to-video, video-to-video, editText-to-video, image-to-video, video-to-video, edit, extend

Built for real workflows

MiniMax H3 is a general-purpose, omni-modal video model: unified understanding across text, images, video, and audio, with generation and editing running through the same architecture. It's built for production work that starts from a prompt, a reference, or a clip you already have.

Product and brand footage

Swap an object, relight a scene, or add a prop to existing footage without a reshoot.

Reference-driven character performance

Carry a subject's look, motion, or voice into a brand new scene using a mix of image, video, and audio references.

Storyboarding with keyframes

Pin the first and last frame of a shot and let H3 generate the motion in between.

Clip continuation

Extend an existing clip or audio segment into a seamless new video, rather than starting the shot over.

Scene relighting and background swaps

Change the time of day or the backdrop while the subject, geometry, and camera path stay put.

Social-ready vertical video

Generate directly at 2K in six formats, from cinematic 21:9 to vertical 9:16, without cropping after the fact.

Get started

One API call, in whichever language you're already using, with the same key and billing as every other model on Runware. To run a variant, change the model to its ID: minimax:h3@max, minimax:h3@max-turbo, minimax:h3@fast, minimax:h3@highspeed, or minimax:h3@lite.

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "<uuid-v4>",
      "model": "minimax:h3@0",
      "positivePrompt": "A slow dolly-in on a lighthouse keeper climbing a spiral staircase at dawn. Sound: wind howling, footsteps echoing on iron stairs.",
      "duration": 10,
      "seed": 1976
    }
  ]'

Response

[
  {
    "taskType": "videoInference",
    "taskUUID": "<uuid-v4>",
    "status": "success",
    "videoUUID": "3fa4c8e2-9b71-4e05-8c3a-1d2f6a9b7e40",
    "videoURL": "https://vm.runware.ai/video/ws/2/vi/3fa4c8e2-9b71-4e05-8c3a-1d2f6a9b7e40.mp4",
    "cost": 1.3,
    "seed": 1976
  }
]

MiniMax H3 Image is now on Runware

MiniMax's image model generates images from text and edits them with up to 16 reference images, at 1K, 2K, or 4K, or at custom sizes up to 20 megapixels. It runs on the same key and billing as MiniMax H3.

Jellyfish posters, sketches, and product designs spread across a wooden floor

Frequently asked questions

What is MiniMax H3?

MiniMax's video model, hosted on Runware's API. It generates and edits video from text, image, video, and audio inputs, with native synchronized stereo audio in every output rather than a separate dubbing pass.

What generation modes does it support?

Text to video, first-frame and first-and-last-frame keyframe generation, reference-driven consistency (carrying a character, product, or style from reference images), motion and camera transfer from a reference video, and instruction-based editing of an existing clip. All of it runs through the same Runware integration.

Does it generate audio automatically?

Yes, by default. Every clip comes back with synced stereo sound generated alongside the picture, and it responds to a written Sound: clause the same way it responds to a description of the image: ambience, effects timed to specific frames, music, and lip-synced dialogue.

What resolutions and durations are supported?

MiniMax H3 makes clips from 4 to 15 seconds, at either 768p or 1440p (marketed as 2K), across six aspect ratios per tier: 16:9, 21:9, 4:3, 1:1, 3:4, and 9:16. The faster variants run at 480p and 768p in the same six aspect ratios, for 4 to 15 seconds, or 5 to 15 seconds on H3 Max and H3 Max Turbo.

What are the MiniMax H3 variants?

Five faster, lower-cost versions of MiniMax H3, each at 480p and 768p with synchronized audio: H3 Max, H3 Max Turbo, H3 Fast, H3 Highspeed, and H3 Lite. H3 Max is tuned for faster generation with strong prompt adherence, and H3 Max Turbo is a distilled H3 Max for quick iteration from text or frames. H3 Fast suits previews and batch runs that need reference images, video, and audio. H3 Highspeed balances quality and speed for everyday content, and H3 Lite is built for low-cost drafts and variations.

Which MiniMax H3 model should I use?

MiniMax H3 for finished work at 2K, and for editing an existing clip, which only it does. For faster, cheaper generation at 480p or 768p, pick a variant: H3 Lite or H3 Max Turbo for drafts, H3 Fast or H3 Highspeed when you need reference images, video, and audio, and H3 Max for stronger prompt adherence. All six take the same request format, so switching is a change of model ID, as long as the request uses only what the new model supports: no 2K output or clip editing on the variants, and no reference inputs on H3 Max Turbo. H3 Max has its own guide to generating video in the docs.

How many reference inputs can I pass in one generation?

Up to 9 reference images, 3 reference videos, and 3 reference audio tracks in a single call, with combined video and audio duration capped at 15 seconds. First-and-last-frame mode is a separate path and can't be combined with reference images, video, or audio in the same request.

How does editing work?

Pass an existing clip in referenceVideos, then write the prompt as an instruction rather than a scene description, naming the change and what should stay untouched. MiniMax H3 returns the same shot with just that change applied, holding the subject, motion, and camera path.

Do I need separate integrations for each mode?

No. Generation, reference-driven composition, motion transfer, and editing all run through the same Runware API endpoint, under the same key and billing as every other model on the platform.

How much does MiniMax H3 cost?

Pricing is metered per second of output. $0.08 per second at 768p, or $0.13 per second at 1440p (2K). A 5-second clip costs $0.40 at 768p or $0.65 at 2K.

How much do the MiniMax H3 variants cost?

Each is metered per second of output, at 480p and 768p respectively: $0.05 and $0.08 on H3 Max, $0.025 and $0.04 on H3 Max Turbo, $0.046 and $0.08 on H3 Fast, $0.04 and $0.06 on H3 Highspeed, and $0.025 and $0.04 on H3 Lite. Reference images cost $0.03 each beyond the first 5 on H3 Highspeed and $0.02 each beyond the first 5 on H3 Lite. H3 Fast, H3 Highspeed, and H3 Lite bill uploaded reference video per second, at their output rate.

Does it cost extra to pass in reference images, video, or audio?

Reference images: the first 5 are free, then $0.04 each. Reference video: $0.13 per second uploaded. Reference audio: free.

Is MiniMax H3 available now?

Yes. MiniMax H3 and its five variants, H3 Max, H3 Max Turbo, H3 Fast, H3 Highspeed, and H3 Lite, are live on Runware, available through the API and Playground under the same key and billing as every other model on the platform.

Does MiniMax H3 generate images?

Images come from MiniMax H3 Image, a separate model on Runware. It generates images from text and edits them with up to 16 reference images, under the same key and billing. See the MiniMax H3 Image page for details.

Already integrated with MiniMax H3 somewhere else?

Our Migration Skills and docs help map an existing integration over. It's the same request shape as every other Runware model, so the change is usually just field names and auth. See the migration guide for details.

Talk to us about MiniMax H3

Have questions about MiniMax H3? Chat to our team about enterprise usage, including volume discounts and dedicated RPM, and we'll follow up shortly.