MiniMax H3 API

What it adds over Wan 2.2

Two things: sound and LoRA depth. H3 is omni-modal, so the clip ships with a generated soundtrack that follows the prompt — ambience, footsteps, voice-like murmur, music cues. And it loads LoRAs the same way the WAN worker does, by URL, so a motion or style LoRA trained for H3 is one request parameter away.

The canvas follows your start frame (768 px short edge, capped at 768×1344), frames are snapped to the model's native grid, and the 8-step turbo path is on by default. There is no negative prompt on this model — describe what you want to see and hear.

• Prompt both motion and sound: "she laughs and turns toward the window, rain on the glass, soft piano"
• LoRAs must be MiniMax H3 LoRAs — WAN or SDXL files will fail to load
• A cold worker loads ~45 GB of weights before the first clip; use a webhook, not a blocking request

Generate a clip

curl https://api.fastgencloud.com/v1/videos/generations \
  -H "Authorization: Bearer $FASTGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "model": "minimax-h3",
      "image": "https://example.com/still.jpg",
      "prompt": "she turns and smiles, wind in her hair, distant traffic and birdsong",
      "duration": 8,
      "loras": [{ "url": "https://huggingface.co/…/my-h3-lora.safetensors", "strength": 0.9 }]
    }'

Parameters

model

minimax-h3

image

start frame — https URL or data URI; the canvas follows its aspect ratio

prompt

optional motion + sound description, up to 2000 characters

duration

5, 8 or 10 seconds

loras

up to 3 MiniMax H3 LoRAs by URL ({ url, strength })

seed

optional integer for reproducible output

webhook_url

strongly recommended for video

Unfiltered and uncensored

This endpoint applies no content filter. Your prompt, negative prompt, reference images and LoRAs reach the model verbatim, nothing is rewritten or injected, and outputs are not scanned or blocked. What your product allows is your call.

The one automatic rule is the minor-safety floor: a prompt that pairs a minor or age reference with sexual content is refused for that request only. Nothing else is filtered and the account is never touched by moderation.

How the API works

Every generation is a job. The POST returns 202 with a job id immediately, so your request thread is never blocked on a GPU. You then either poll GET /v1/jobs/{id} or register a webhook_url and receive a signed callback when the job finishes.

Send an Idempotency-Key header and a retried request replays the original job instead of charging you twice. Your wallet is debited when the job is submitted and refunded automatically if it fails.

curl https://api.fastgencloud.com/v1/jobs/$JOB_ID \
  -H "Authorization: Bearer $FASTGEN_API_KEY"

Pricing and billing

Pricing is prepaid and usage-based: you top the wallet up with a card and each job debits it. There is no subscription, no per-seat cost, and no monthly minimum.

Images above 1 megapixel are billed pro rata by pixel count — a 1.5 MP image costs 1.5x the listed rate. Failed jobs are refunded in full, automatically.