SYLIN

MiniMax H3

by MiniMax · Model · Video

strong

Best open video model on current evidence, with a restrictive license

#1 video editing (open weights)#1 video generation (open weights)

#1 on Arena video-edit among open weights (1378)

AccessOpen weightsNeeds~54 GB; smallest build ~5 GBDownloadabout 34 GB; running needs about 54 GBLicenseMiniMax Community License

Why we say so

Arena ranks it first among open weights for text-to-video (1456, well ahead of Kandinsky Video Pro 5.0 at 1175) and for video editing. AI Search's reviewer called it the best open video model and found Kandinsky 6.0 below it. Sizes differ by source: weights download at 34 GB, while one source says 54 GB is needed to run it.

Evidence

  1. Nov 17
    Commenters propose a harder physics test: describe the scene without the outcome and see if the model predicts what objects do
  2. Nov 9
    #1 on Arena text-to-video among open weights (1456)
  3. Nov 8
    Described as the best open video model; base of P-H3 Image
  4. Sep 22
    #1 on Arena text-to-video among open weights (1460)
  5. Sep 20
    Base video model that Meridian and Jing/Dao fine-tune; called the best video model out there
Show all 16
  1. Aug 27
    #1 on Arena video-edit among open weights (1392)
  2. Aug 27
    Used locally on a 12 GB RTX 4070 to lip-sync animation to recorded dialog, which Gemini Omni cannot do
  3. Aug 23
    ComfyUI setup uses Comfy-Org/MiniMax-H3 files: ~34 GB diffusion model, separate video and audio VAEs, text encoder (AWQ variant available)
  4. Aug 12
    Open-source video generator with built-in audio, runnable locally; reviewer calls it the best open video model so far
  5. Aug 12
    MiniMax community license allows commercial use under US$20M yearly revenue; EU, UK, Korea and US excluded unless a separate license is approved
  6. Aug 12
    Took #1 on Arena video-edit
  7. Aug 12
    Took #1 on Arena video-edit among open weights
  8. Aug 12
    H3 can generate still images with a community workflow, but reviewer says Ideogram is much better and wouldn't use H3 as an image model
  9. Aug 5
    Described as an open-source Gemini Omni for multimodal reference and editing
  10. Aug 3
    Took #1 on Arena text-to-video among open weights
  11. Aug 3
    Music video from an uploaded ACE-Step track: MiniMax H3 kept the song exact, synced cuts to beat with good typography; Seedance output was weird

Cost to run

Alternatives

SeedanceFight scene from 3D reference video plus character images: Seedance followed motion with consistent characters; MiniMax H3 ignored the reference, with noise and artifacts
Kandinsky Video Pro#2 on Arena text-to-video among open weights (1175)
HunyuanVideo#3 on Arena text-to-video among open weights (1168)
LTXReviewer measured LTX 2.5 as more than two times faster than MiniMax H3 on their PC
More MiniMax H3