Skip to content
CozyGPU

MiniMax H3

Open-weight video with sound. Five to fifteen second clips from text or a still image.

Modes
Text to video · Image to video
GPU class
80 GB Ada Pro or 96 GB Blackwell
RunPod pool price
$3.49 to $4.79/h
Typical cost per run
$0.16 to $0.32 per 5 s clip
01

What it is

MiniMax H3 is an open-weight video model that generates picture and audio together. Clips run from five to fifteen seconds on a 768p canvas, from a text prompt or from a single input image that sets the composition and aspect ratio.

Video is the expensive one. CozyGPU runs it on an 80 GB Ada Pro or a 96 GB Blackwell worker, and a five second clip takes two to four minutes of GPU time. The app shows the cost of each job in the library so there are no surprises.

Weights on Hugging Face: Comfy-Org/MiniMax-H3
02

Modes

  • Text to videoAspect presets 16:9, 3:2, 4:3, 1:1 on a fixed 768p canvas.
  • Image to videoOne input image; the canvas follows the image's aspect.
03

Settings exposed

  • Length5 to 15 seconds.
  • Aspect16:9, 3:2, 4:3 or 1:1 (text to video); from the input image (image to video).
  • Output frame rate24 fps native; 30, 50 or 60 by frame interpolation.
  • AudioGenerated with the picture; can be switched off.
  • SeedRandom, or fixed.
  • LoRAsThe built-in Realism People LoRA plus your own, each with a strength slider and trigger words.
  • Custom LoRAsSupported. A built-in Realism People LoRA ships with the worker; your own .safetensors files load by drag and drop and work in both text to video and image to video, up to two per job.
04

Rendered with CozyGPU on the GPU class above, straight from the library. Costs are what RunPod billed for each job.

  • Image to video from the silk render: the silk flows and ripples slowly like a wave$0.14
  • Image to video from the smoke render: the wisp curls and drifts upward slowly$0.09

Other models