stimi/minimax-h3-lite-api

MiniMax H3 (open weights) video API on apimodels.app

This page documents the lower-priced MiniMax H3 video line served by apimodels.app, built on the open MiniMax H3 weights. This repository contains no model weights. It describes how to call the model through the apimodels.app API, with real examples, prices and limits.

Not MiniMax's official H3 Lite. MiniMax's own H3 Lite is a separate product. This line has been offered by apimodels.app since August 2026 under the model id minimax-h3-lite: it accepts reference images, reference audio and first/last frames, but not reference videos, and it takes minutes per clip.

• Model page: apimodels.app/models/minimax-h3-lite
• API reference: apimodels.app/docs/minimax-h3-lite
• Get an API key: apimodels.app/console/api-keys
• Run it in a browser: replicate.com/stimqq/minimax-h3-lite (uses your own apimodels.app key)

Examples

Both clips below were generated by this model exactly as shown: same prompt, same inputs, no editing.

1. Four reference stills, one 12-second take

768p · 12 s · 16:9 · 4 reference images · $0.24

Reference images (Picture 1 to 4: face, full look, location, prop):

Prompt:

Picture 1 is the woman's face — lock her identity exactly: short black bob with a grey streak, the small scar above her left eyebrow. Picture 2 is her full look — olive military parka, red scarf, wet hair. Picture 3 is the location: the rain-soaked harbor pier at dusk with the single sodium lamp and fog. Picture 4 is the brass pocket watch with the cracked face.
12-second single continuous take, 16:9, photoreal, 35mm anamorphic, rain, native ambient audio (rain on concrete, distant foghorn, the watch's faint tick that stops).
0-3s: Low tracking shot gliding over wet concrete toward her boots; she crouches, picks the watch up from a puddle, water runs off it. 3-6s: Camera rises to a tight close-up as she turns it over in her hands; her face is guarded, jaw tight, breathing controlled. 6-9s: She hears the tick stop — a flicker of hope collapses; her lower lip trembles, eyes fill but she refuses to blink, a single tear breaks and runs down with the rain. 9-12s: She closes her fist around the watch, exhales, lifts her chin toward the fog with hardened resolve; slow push-in ends on her eyes catching the sodium light. Rack focus from watch to face at 6s. No dialogue. No text.

2. Lip-synced monologue from a voice track

768p · 15 s · 16:9 · 2 reference images + 1 audio clip · $0.30

Reference images (Picture 1: face, Picture 2: wardrobe and pier) and Audio 1 (the voice line):

Prompt:

Picture 1 locks the woman's face (short black bob with grey streak, scar above left eyebrow). Picture 2 locks her wardrobe (olive military parka, red scarf) and the rain-soaked pier at dusk. Audio 1 is her voice — she speaks these exact words on camera, lip-synced, addressing someone just behind the lens.
15-second single take, 16:9, photoreal, handheld with slight breathing sway, rain and fog, sodium lamp key light, native audio = Audio 1 plus rain ambience only.
Performance: she starts almost calm with a bitter half-smile; on "I waited on this pier every night" her voice cracks and her eyes glisten; on "and now it has stopped" she looks down at the pocket watch in her palm and swallows hard; on "Just like you did" she looks straight into the lens, tears finally falling, then she turns away into the fog. Medium close-up drifting to close-up. No subtitles.

The character is fictional and AI-generated.

What it does

One model id, four ways to steer it:

• Text to video — a prompt alone.
• Reference images — up to 9 images of a character, outfit, product or scene to keep consistent. Refer to them in the prompt as Picture 1, Picture 2, …
• Reference audio — up to 3 clips (a voice line for lip-sync, or music to move to), together with at least one reference image. Refer to them as Audio 1, …
• First / last frame — anchor the opening frame, or both the opening and closing frames.

Output: 480p or 768p, any whole length from 1 to 15 seconds, 16:9, 9:16 or adaptive, with synchronized audio.

Pricing

Billed per second of output from your apimodels.app balance. Failed generations are not charged.

Resolution	Per second	5 s clip	15 s clip
480p	$0.01	$0.05	$0.15
768p	$0.02	$0.10	$0.30

Quick start

# 1. create the task
curl https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer $APIMODELS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3-lite",
    "prompt": "Picture 1 is the woman. She turns toward the camera on a rainy pier at dusk.",
    "images": ["https://example.com/face.jpg"],
    "duration": 5,
    "resolution": "768p",
    "ratio": "16:9"
  }'

# 2. poll until data.state is "completed", then read data.resultUrls[0]
curl "https://api.apimodels.app/v1/video/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer $APIMODELS_API_KEY"

Images and audio can be public URLs or base64 data URIs. Fields: prompt, images (≤ 9), audio_list (≤ 3), first_frame_url, last_frame_url, duration (1–15), resolution (480p / 768p), ratio (16:9 / 9:16 / adaptive), optional callback_url. Python and Node examples: API reference.

Limits

• Not a fast model: 480p · 3 s takes about 1–2 minutes; 768p · 10 s about 6 minutes; 15 s clips 8–11 minutes, longer when the queue is busy.
• No reference video on this line.
• First/last-frame mode cannot be combined with reference images or audio; a last frame needs a first frame.
• Success is not 100%. Safety filters can reject a prompt or an image, and the model can be temporarily unavailable. Nothing is charged in either case.
• Result files are kept for 30 days; download what you want to keep.

Credits and license

MiniMax H3 is developed by MiniMax and released under the MiniMax H3 Community License Agreement. apimodels.app is an independent API platform and is not affiliated with MiniMax, Hugging Face or Replicate.