MiniMax H3 API: H3 and H3 Max video at native 768p with stereo audio

MiniMax H3 and MiniMax H3 Max are available through the Sume API as minimax-h3 and minimax-h3-max on POST /v1/videos. Both make 5–15 second clips at native 480p or 768p with native stereo audio and take image, video, and audio references; H3 Max adds 1080p, a latent refinement from native 768p.

The facts below come from Sume's Video generation and Video Router docs and from the model catalog the API serves, read on 2026-09-26. The job flow is the same for every catalog model; An OpenRouter-compatible video API walks through it.

What is the difference between H3 and H3 Max?

The two ids share a duration range, native resolutions, modes, and reference types. They differ above 768p.

Why send 768p instead of 720p?

Because 768p is the native tier. The catalog calls it first-class and says not to treat it as 720p, and 720p is not in either model's list, so it is refused. Send resolution: "768p" or "480p".

On minimax-h3 you can also ask for 2K or 4K. Both are upscales of the native 768p output, priced if requested, and they do not appear in the model's supported_resolutions. On minimax-h3-max, 1080p is the only tier above 768p.

Can I turn off the audio?

No. Both models always produce native stereo audio, so omit generate_audio. Sending generate_audio: false to POST /v1/videos returns 400 unsupported_capability with a message to omit the field. Neither model takes bitrate_mode.

How many references can I send?

Put references in input_references. For image-to-video, send frame_images with a first_frame and an optional last_frame instead; if you send both fields, frame_images wins. The reference limits are the same on both models:

• Up to 9 images.
• Up to 3 videos, each 2–15 seconds, 15 seconds combined.
• Up to 3 audio clips, each 2–15 seconds, 15 seconds combined.
• No more than 12 references in total.
• Audio cannot be the only reference; pair it with an image or a video.

curl -X POST "https://api.sume.com/v1/videos" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: h3-max-ref-001" \
  -d '{
    "model": "minimax-h3-max",
    "prompt": "A barista pours latte art at a sunny cafe counter",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/barista.png" } },
      { "type": "audio_url", "audio_url": { "url": "https://example.com/cafe.mp3" } }
    ],
    "resolution": "768p",
    "aspect_ratio": "9:16",
    "duration": 10
  }'

How are MiniMax H3 and H3 Max billed?

Per output second, at a rate set by resolution. In pricing_skus, minimax-h3 has a rate for 480p and 768p, and minimax-h3-max adds 1080p. Two minimax-h3 charges are not listed there: the 2K and 4K upscales, and a charge for each reference image after the first five in reference-to-video. The model's constraints in GET /v1/video-router/models/minimax-h3 state both. Both models, H3 Max included, are billed at the provider's list price × 1.25 like every model, plus a 5.5% agent fee by default; List video generation models via API explains the SKUs.

Is there a MiniMax lip-sync endpoint?

Yes, as a separate product. MiniMax H3 Max Lip Sync (minimax/h3-max/lip-sync) runs at POST /v1/minimax/h3-max/lip-sync and takes the same still-plus-audio body as VEED Fabric 1.0, with 5–14.8 seconds of audio, billed at list × 1.25, per the Models overview. The lip sync API guide covers it.