MiniMax H3 Lite
⧉MiniMax H3 Lite is the same MiniMax H3 model run with fewer inference steps: faster and cheaper, with slightly lower quality, made for drafts in bulk and prompt tests. On KovAPI it is minimax-h3-lite: 480P or 768P, any length from 4 to 15 seconds, billed per second; for a first frame, a reference video or 1080P, switch to Standard.
- 480P from $0.005/s
- Any length 4–15 s
- Up to 5 reference images
- Failures not billed
Live Test · Playground
Try out MiniMax H3 Lite right here (available after login).
Input
Output
videoPicture and sound in one pass, with control over the opening, the ending, the cast and the motion
These are the MiniMax H3 family's capabilities, mostly shown with Standard clips; Lite does text to video and reference images only (no first frame, no reference audio or video), and "Lite: a turntable shot of a running shoe" was made with Lite. "Full example" opens the original, its prompt and settings.
Shots, lines and sound follow the prompt's timeline
Written shot by shot in the official format, the cut times, the dialogue, the ambience and the music follow the prompt. The bakery clip cuts to the bread knife at 5 seconds, the baker says his line word for word, and the acoustic guitar is the one the prompt asked for.
Fix the opening, or both the opening and the ending
One image as the first frame and the clip starts from it; two images as first and last frame and the model fills in the motion between. The umbrella clip starts on the first picture with the umbrella furled and ends on the second with it open.
People and products from your images, combined the way you write
Reference mode takes up to 9 images. The eyewear ad uses three, the models, their faces and the glasses; written in the reference guide's structure, the models put the glasses on, walk the runway and the clip cuts to a close-up.
Borrow the motion and camera of a reference video
Left, the reference video; right, the result: a cartoon fox dances the same moves as the person. A reference video can also give only the camera move, while the reference image and the prompt decide what is in the shot. Reference seconds are billed too.
Draft at 480P, deliver at 1080P
480P / 768P / 1080P, eight aspect ratios, any length from 4 to 15 s. Check framing and motion at 480P or on Lite, then render the final clip at 1080P; changing tier renders a new clip, not the same one, and the same-prompt comparison below shows the difference.
MiniMax H3 Lite at a glance
- Model ID
minimax-h3·minimax-h3-lite- Endpoint
POST https://kovapi.com/v1/videos/generations→GET https://kovapi.com/v1/videos/jobs/{id}- Length
- 4–15 s, whole seconds
- Resolution
- 480P / 768P; 768P when omitted; 720P is read as 768P
- Aspect ratio
- 16:9 / 9:16 / 1:1 / 4:3 / 3:4 / 3:2 / 2:3 / 21:9
- References
- up to 5 reference images (references only, never a first frame); no reference video or audio
- Audio
- every clip has a native stereo track (dialogue, effects, music); it cannot be turned off
- Prompt
- ≤ 5,000 characters
- Lite per second
- 480P $0.005/s · 768P $0.0125/s
- Results kept
- 7 days
Modes and API fields
Both modes share model (minimax-h3-lite), prompt, duration, resolution and size. Images are always references, so image_role is not needed.
| Mode | Send | You get | Limits |
|---|---|---|---|
| Text to video | Just prompt and the generation settings. | Picture and sound from the text; write dialogue, effects and music into the prompt. | Any whole number of seconds from 4 to 15; 768P without resolution. |
| Reference images | 1–5 images in images. | The people, products or places in them appear in a new shot, never as the first frame. | Up to 5; image_role: first_frame is refused. |
Combination rules
- Lite has no first frame or first + last frame: images are always references, and
image_role: first_frameis refused at once and not billed. To open a clip on a given image, use Standard. - Lite takes no reference video or reference audio; a request with either is refused and not billed. Use Standard for those.
- Only 480P and 768P:
1080Pis refused and720Pis read as 768P. Every clip has sound; for no music writenon_diegetic_music: N/Ain the prompt.
MiniMax H3 prompts: follow the official structure
Based on MiniMax's own Video Prompt Writing Guide (base, reference mode). MiniMax says prompt structure matters more on H3 than on most models: write the picture shot by shot, put dialogue in tags, and give ambience and music a section each. The habits below come from that guide and the examples are prompts this page's clips were made with; points marked "tested" are what our own renders showed.
Habits that work
- Write three sections
integrated_multimodal_descriptionfor picture, action, shots and dialogue;overall_soundscapefor ambience and action sounds in 1–4 sentences;non_diegetic_musicfor music only the audience hears, orN/A. Tested: the same bakery scene as a one-liner still got its line and its cut, but no music and the wrong sounds; in three sections the guitar, the doorbell and the bread knife all came out as written.Bakery, opening words: “integrated_multimodal_description: [Shot 1] Live-action, cinematic, …”
- One block per shot, with the cut time
[Shot 1]has no time; later shots read[Shot 2] At 00:05.000, the camera cuts to…. In the multi-shot clips on this page the cuts landed close to the times written.Bakery: “[Shot 2] At 00:05.000, the camera cuts to a close-up of steam rising from the sliced bread”
- Camera = type + amplitude + speed
Push in, pan, truck, arc shot, tracking shot, plus with small amplitude or at slow speed, written as part of the action rather than tags stacked at the end.
Bakery: “The camera pushes in with small amplitude at slow speed as the middle-aged baker…”
- Dialogue in its own tags
Number speakers (S1), (S2), describe age and voice first, then write the line as
<d>[English] …</d>, word for word and untranslated. Chinese lines use[Chinese].Train:
the quiet, breathy young woman (S1) says: <d>[English] I get off at the next station.</d> - On-screen words in quotes
Put the words of a sign or banner in English double quotes exactly as they should appear. Tested: “营业中” came out right, while signs not written into the prompt came out as gibberish.
Noodle stall: “A red neon sign reading "营业中" glows above the stall”
- Chinese dialogue comes with captions
Tested: both renders of a Chinese line burned Chinese captions into the bottom of the frame, and asking for no captions did not remove them; the English lines had none. For a clean frame use English dialogue or fix it in post.
Noodle stall, last sentence (it did not work): “No subtitles or captions appear on screen.”
- Frames: alignment line and path
A first-frame task opens with the alignment line; first + last says which second the second picture lands on, and the body describes the motion between them instead of describing both pictures again.
Umbrella: “Picture 2 (from Shot 1) aligns with the 8.00-second mark of the target video”
- Name every reference
Label each input
<Subject 1>,<Picture 1>and say what it does in the clip. Tested: the one-line style of MiniMax's highlights page (“Image 1: …, Image 2: …”) turned three reference images into three nearly still shots; rewritten in the reference-guide structure, the models put the glasses on and walked.Eyewear:
<Subject 3> is the futuristic silver sunglasses in <Picture 3>…
Anatomy of a long prompt
- 1Picture section
Starts with
integrated_multimodal_description:, then one block per shot.Original: “integrated_multimodal_description: [Shot 1] …”
- 2Style and framing
The first shot opens with style and shot size.
Original: “Live-action, cinematic, a medium-wide shot frames a baker opening the shutters…”
- 3Camera and action
Camera type, amplitude and speed in the same sentence as the action.
Original: “The camera pushes in with small amplitude at slow speed as the middle-aged baker… places a fresh loaf on the wooden counter”
- 4Dialogue
Speaker number, voice and the tagged line.
Original:
with a calm, slightly raspy voice (S1) … says: <d>[English] First batch of the morning.</d>- 5Cut
A new shot gives its cut time and the new picture.
Original: “[Shot 2] At 00:05.000, the camera cuts to a close-up of steam rising from the sliced bread”
- 6Two sound sections
One for ambience, one for music.
Original: “overall_soundscape: Wooden shutters scrape open… non_diegetic_music: A soft acoustic-guitar pattern…”
A template to edit
integrated_multimodal_description: [Shot 1] [style, e.g. Live-action, cinematic], [shot size] frames [subject and setting]. The camera [move] with [small / large] amplitude at [slow / fast] speed as [action]. The [speaker's age and voice] (S1) says: <d>[English] [line]</d> [Shot 2] At 00:0X.000, the camera cuts to [next picture]. overall_soundscape: [ambience and action sounds, 1–4 sentences] non_diegetic_music: [instruments, tempo and dynamics; N/A for no music]
MiniMax H3 pricing
Billed per second, each tier at its own rate; the playground shows the price of a request before you submit.
| Resolution | Standard | Lite |
|---|---|---|
| 480P | $0.015/s | $0.005/s |
| 768P | $0.025/sMiniMax list price $0.08 · 68% below list | $0.0125/s |
| 1080P | $0.036/s | — |
- One clip costs the tier's rate × (output seconds + reference-video seconds); reference images and audio are free. Drag "Reference video" in the calculator to price a request with one.
- MiniMax's own price list has H3 at 768P and 2K only, so 768P is the one tier with an official price beside ours.
- The amount is held at submit and settled as is on success; failed jobs are refunded in full. For cheaper drafts in bulk, see MiniMax H3 Lite.
Cost calculator
One prompt, five tiers
Hover over the clips to play them together; on a phone, tap "Play together". Each clip is a separate render, so they are not identical: compare sharpness and detail.
Latte art: three H3 tiers and two Lite tiers
The same prompt, 5 seconds at 16:9, rendered once each on H3 at 1080P / 768P / 480P and on Lite at 768P / 480P. Check framing and motion at 480P or on Lite, then render the final clip at 1080P.
Sharpest: the latte pattern, the jug's reflections and the cup rim are all crisp.
A little less detail than 1080P, still clean.
Framing and motion are clear; soft when enlarged. Good for drafts.
Fast and cheap; the latte pattern is clearly weaker.
Cheapest; only for checking framing and motion.
Original prompt
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a close shot of a barista's hands pouring steamed milk into a cup of espresso on a wooden counter by a sunny window. The milk forms a heart in the crema while steam curls upward, and the camera pushes in with small amplitude at slow speed toward the cup. overall_soundscape: Milk streams into the cup, a steam wand hisses softly in the background and cafe chatter murmurs. non_diegetic_music: N/A
MiniMax H3 or another video model?
Lengths, resolutions, reference limits and starting prices in the table are live.
| Model | Length (s) | Resolution | Refs image / video / audio | Price | Good for |
|---|---|---|---|---|---|
MiniMax H3 LiteThis pageminimax-h3-lite |
4–15 | 480P / 768P | 5 / 0 / 0 | from $0.005/s | Bulk drafts and prompt tests: the same H3 with fewer inference steps, faster and cheaper with slightly lower quality; text and reference images only. |
MiniMax H3minimax-h3 |
4–15 | 480P / 768P / 1080P | 9 / 3 / 3 | from $0.015/s | 1080P clips with sound, first + last frames, or reference images with a reference video; cheap 480P drafts too. |
Wan 3.0wan3.0-video |
4–30 | 480P / 720P / 1080P | 10 / 5 / 5 | from $0.04/s | Clips longer than 15 s (up to 30), or video from a document or a web page. |
Seedance 2.5seedance-2.5 |
4–30 | 480P / 720P / 1080P | 30 / 10 / 10 | from $0.10/s | Lots of reference material, or story clips up to 30 s. |
Seedance 2.0 Fastseedance-2.0-fast |
5–15 | 480P / 720P | 9 / 3 / 3 | from $0.046/s | Fast, low-cost Seedance clips. |
Gemini Omni 1.2 Flashgemini-omni-1.2-flash |
3–15 | 720P / 1080P / 4K | 3 / 0 / 0 | from $0.035/s | Editing an existing video with one sentence, 4K, or exact on-screen text. |
Veo 3.1veo-3.1 |
8 | 720P / 1080P | 3 / 0 / 0 | $0.20/clip | Fixed 8-second clips billed per clip. |
Grok Imagine Video 1.5grok-imagine-video-1.5 |
4–15 | 480P / 720P | 7 / 0 / 0 | $0.30/clip | 480P / 720P clips billed per clip, with up to 7 reference images. |
If you already have clips from another model, run the same inputs and acceptance checks on both and compare picture, motion, sound, render time and real cost before switching. For H3 versus Lite, see the same-prompt comparison above.
How to generate video with MiniMax H3 Lite on KovAPI
-
1
Sign up and create an API key
Sign in to KovAPI and create a key in the console; one key calls every model on the site.
Sign up for KovAPI -
2
Try prompts at the top of this page
Pick length, resolution and aspect ratio, add reference images, and see the live price before submitting.
Open the playground -
3
Submit an async job
Send
Read the MiniMax H3 Lite API docsminimax-h3-litetoPOST /v1/videos/generationsand keep the job id from the HTTP 202 reply; full request examples are in the docs. -
4
Poll the result or take a webhook
Poll the job id, or set
Read the webhook docscallback_urlwhen submitting. On success download the MP4 fromdata[0].url; clips are kept 7 days, so save them in time.
Why use MiniMax H3 Lite on KovAPI
Cheap drafts
Lite costs $0.005/s at 480P and $0.0125/s at 768P. Try framing and motion for a prompt on Lite in bulk, then switch the chosen ones to Standard at 1080P. The playground shows the price of a request before you submit.
Failures are free and retried
If a job fails, the amount held is refunded in full; when the upstream is busy or errors out, the system retries for you, so you need no retry logic of your own.
No concurrency cap
No account has a limit on parallel jobs, so submit batches straight away; at peak times jobs queue and the API returns queue_position and eta_seconds, with a webhook when they finish.
One API for every video model
Submitting, polling and webhooks work the same for every video model on the site; change model to send the same prompt to Standard, Wan, Seedance and others.
Easy sign-up and payment
Sign up with email, Google or GitHub; top up by WeChat, card, PayPal or USDT; pay as you go with no monthly fee, and the same key also calls the chat and image models on the site.
How MiniMax's video models evolved
Dates from MiniMax's own release notes; what KovAPI offers today is in the mode table above.
- 2025-02-11
T2V-01-Director · camera control
Camera instructions written in the prompt, and the model moves the camera as told; H3's prompt guide keeps this way of writing camera moves.
- 2025-06-18
Hailuo 02 · 1080P
The second-generation video model, with 1080P and clips up to 10 seconds.
- 2025-10-28
Hailuo 2.3 / 2.3 Fast
Better body movement, facial expression, physical realism and prompt adherence; the Fast version is image-to-video only.
- 2026-07-31
MiniMax H3 · third generationCurrent
Text, images, video and audio understood together, picture and stereo sound in one pass; first frame, first + last frame, reference images and reference video in one model. KovAPI offers 480P / 768P / 1080P.
MiniMax H3 API FAQ
What is MiniMax H3?
minimax-h3, with text to video, first frame, first + last frame, reference images and reference video.How long can a clip be?
duration. A clip runs a fraction of a second longer than requested (a 5-second job is about 5.2 s) and is billed for the seconds requested.Which resolution should I pick, and how big is it?
resolution is omitted. At 16:9, 480P is 864×480, 768P about 1344×768 and 1080P about 1920×1080; a job can differ by a few dozen pixels.Can I send 720P?
How do I switch between frames and references?
image_role, one image is the first frame and two are the first and last frame. With image_role: reference they are reference images, up to 9: their people, products and places appear in the shot instead of being used as frames. Reference video or audio needs reference mode.What can a reference video do, and how is it billed?
Does reference audio help, and is it billed?
What formats and sizes can the references be?
Can I turn the sound off?
audio: false is refused. For no music write non_diegetic_music: N/A; for a silent file, remove the track after downloading.Can the dialogue be in Chinese or other languages?
<d>[Chinese] …</d>; our Chinese line came out word for word, but Chinese captions appeared at the bottom of the frame and asking for none did not remove them. MiniMax says dialogue is stable in 11 languages, including Chinese, English, Japanese, Korean, Spanish and Russian. Writing the rest of the prompt in English, as the guide does, is safer.Can I use photos of real people?
How long does a render take?
callback_url and wait for the call.Am I charged when generation fails?
How long are clips kept?
H3 or H3 Lite?
H3, Wan 3.0 or Seedance 2.5?
Where do the examples on this page come from?
Related Models
Explore other models you can integrate.
MiniMax H3
480P / 768P / 1080P · first & last frame · omni reference · native audio
Wan 3.0
One endpoint, five modes · up to 30 seconds
Seedance 2.0 Mini
Entry tier · lowest cost
Grok Imagine Video 1.5
Billed per clip · up to 15 seconds
