SKILL.md
MiniMax H3
MiniMax H3 is the Hailuo 3.0 family. The first-party API identifier is MiniMax-H3; fal.ai exposes the family as hailuo-03, and Runway uses hailuo3. Do not substitute one provider's identifier into another API.
Choose a route
| Route | Tool call | Execution |
|---|---|---|
| MiniMax direct | minimax_video, model: "MiniMax-H3" |
Hosted first-party v2 API; global or mainland-China region |
| fal.ai | minimaxfalvideo |
Hosted gateway; T2V, I2V, reference-to-video |
| Runway | runway_video, model: "hailuo3" |
Hosted; 768P or 2K, 5–15 seconds |
| ComfyUI Partner Node | comfyuivideo, modelfamily: "minimaxh3api" |
Hosted and billed in Comfy credits |
| ComfyUI open weights | comfyuivideo, modelfamily: "minimaxh3local" |
Local GPU with official workflow and model stack |
For local ComfyUI, export the official workflow in API format and pass workflowjson or workflowpath plus output_node. OpenMontage reports the required diffusion model, Qwen3-VL text encoder, video VAE, and audio VAE; it does not silently download large weights.
Operations and prompting
- Text-to-video: concrete ratio; do not use
adaptivewithout visual input. - Image-to-video: provide a first frame.
- First/last-frame: provide both images and describe the motion between them.
- Reference-to-video: images, videos, and audio can be combined. Audio needs at
least one visual reference.
Write prompts as subject + action + camera path + environment + lighting + audio intent. MiniMax responds well to explicit camera direction. Keep the requested motion achievable within 4–15 seconds and inspect native audio as carefully as the image track.
Provider differences
The direct MiniMax v2 route currently outputs 2K and supports 4–15 seconds. Runway's Hailuo 3.0 route supports 768P/2K and documents 5–15 seconds. Partner Nodes require network access and credits. Only the open-weight ComfyUI route is local/offline after all models are installed.