CivArchive
    MiniMax H3 - H3 (Comfy)
    NSFW

    Hailuo H3 is the newest video generation model from MiniMax, the Shanghai AI lab behind the Hailuo video app and the MiniMax text and audio model families. H3 generates native 2K video with synchronized audio in a single pass, at clip lengths up to 15 seconds.


    Developed by MiniMax. All credit for the model goes to the MiniMax team. You can use H3 two ways on Civitai: generate on-site against the hosted MiniMax API (the H3 (Api) version), or download the open weights from the H3 (Comfy) version on this page and run it yourself. The ComfyUI-format weights are also available from MiniMax and Comfy-Org on Hugging Face.

    Background

    MiniMax has shipped video models under the Hailuo name since 2024, moving from the I2V-01 series through Hailuo 02 and Hailuo 2.3. Those earlier models topped out around 1080p with fixed 6 or 10 second clips and no native sound. H3 is the step past that: it raises the ceiling to 2K, opens the clip length to any integer from 4 to 15 seconds, and folds audio generation into the same pass as the picture rather than bolting it on afterward.

    Capabilities

    • Native 2K output (2560x1440 at 16:9)
    • Clips from 4 to 15 seconds, in one second increments
    • Synchronized audio generated with the video - dialogue, effects, and ambience timed to what is on screen
    • Text-to-video, image-to-video, and first frame plus last frame control
    • Omni-reference conditioning: up to 9 reference images, 3 reference video clips, and 3 reference audio clips
    • Multi-shot sequences and instruction-based editing

    On Civitai

    H3 is wired into the video generator across four workflows: Create Video (text only), Image to Video (single first frame), First/Last Frame (both endpoints), and Reference to Video (up to 9 reference images). Resolution is fixed at 2K because that is the only output H3 produces. Aspect ratio is selectable at 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 for text prompts; when you supply a frame image, H3 adapts the framing to that image instead.

    Reference video and reference audio inputs are supported by the model but are not exposed in the Civitai form yet.

    Notes

    H3 is a new model and MiniMax has not published a full spec sheet or a stable pricing page for it. Behavior and cost may shift as they finish the rollout. Prompt handling follows the usual MiniMax posture - this is a commercially hosted service and content is subject to the provider's policies, so expect refusals on material their API declines.

    • MiniMax video generation docs
    • Hailuo AI
    • MiniMax
    • MiniMax Hailuo AI Terms of Service

    Description

    FAQ

    Comments (19)

    PastelPastelPastelAug 10, 2026· 29 reactions
    CivitAI

    MINIMAX PLEASE OPENSOURCE YOUR NEXT MODEL ASWELL

    RedditUser9811Aug 11, 2026· 7 reactions

    no they will not, just look at civitai.red under loras thats why

    AnistasiataAug 11, 2026· 1 reaction

    @RedditUser9811 I feel like if you're developing any kind of creative software in 2026 and you don't go into it expecting this use case, you're beyond naive. I'm sure they already knew where users would take their model when they made the decision to open source it.

    unbound_rider1Aug 12, 2026

    No, it's usually bait and switch.

    AnistasiataAug 12, 2026

    @unbound_rider1 Yep. It's human nature, apparently. Civitai creators are doing that now too. Release an awesome Lora, get people to follow, then start dropping the really high-demand stuff behind a forever paywall.

    busy_ad6969Aug 12, 2026· 2 reactions

    Don't fret. If they don't, it'll just be another WAN situation where they fall off the map and some other startup looking to make a name will take their spot.

    bhoppingAug 12, 2026

    @busy_ad6969 exactly, there is so much competition at the moment we shouldn't have to worry about that. I thought we would have to wait a couple years to reach seedance 2 level for open source but here we are (very close). I think this is the first omni reference for local models!

    oaf40Aug 11, 2026· 3 reactions
    CivitAI

    Is there a solution to the low framerate issue in I2VA/FL2VA mode? Minimax generates fake duplicate frames as if the scene was animated at 12-14 FPS despite being targeted at 24 FPS. I'm using the official comfyui workflow and I tried literally everything, from not having anything related to slow/chopped motion and low framerate in the prompt to explicitly specifying "cinematic smooth motion". Nothing helps. Same issue no matter what, with or without Turbo LoRA and/or SAGE Attention. It seems like the model somehow infers low-fps anime style from my projects and I don't know how to make it synthesize smooth videos without duplicate frames.

    contrarianAug 13, 2026· 3 reactions
    CivitAI

    Civitai renames all files to "minimax_h3Comfy", so you have to rename them back to what they were supposed to be named manually, just to tell them apart. So dumb.

    ManiacXAug 14, 2026
    CivitAI

    minimax_h3_video_vae_fp32.safetensors for anyone interested. (probably no visible difference from fp16)

    hamburglerreturns973Aug 15, 2026· 4 reactions
    CivitAI

    Using the on-site generator, my I2V always comes out 16:9 regardless of the image ratio. Is there something I'm missing?

    reikasamaAug 15, 2026

    You can use reference image.

    reikasamaAug 16, 2026· 1 reaction

    When choosing reference image instead you can choose between different formats.

    hamburglerreturns973Aug 16, 2026· 1 reaction

    @reikasama thanks! It works

    reikasamaAug 16, 2026

    @hamburglerreturns973 if you want more coherent results, reference images are better anyway. You can even reference start and ending image.

    reikasamaAug 16, 2026· 6 reactions
    CivitAI

    When using on-site Turbo, the image becomes very pixelated. Looks like a bitrate issue. Base works fine.

    yeti842Aug 26, 2026· 1 reaction
    CivitAI

    WooooW - official account giving something to download? really? what happen? is there a christmas? They censor everything everywhere who was good making sh.. everywhere and this supprise me.

    randombrowser1234Aug 27, 2026
    CivitAI

    Hi and thank you. Here is my issue : when using the on-site trainer to train a real-life character lora on MinimaxH3, the result does not look like my character at all. This is not my first time using the on-site trainer, I have used it a lot with other checkpoints in the past, with a high quality dataset (36 high quality photos), and I know exactly how this works. But for starters, I simply used the default training settings. (3000 steps / 10 epochs) then extended the training to 20 epochs. I tried the models from every epoch from 1 to 20, and none of them bear any accurate facial identity, although the style layers (clothing) seem to be okay and the resulting model has a good sharp quality. What is the issue with the identity though ? Can anyone confirm they had awesomely good results training a lora character model on MinimaxH3 ? If so, did you use the on-site trainer's default settings ? If not, that means I might need to try an alternate dataset. Can a couple of blurry faces out of 36 images dataset wreck the result so much as to make the character look like someone else? I doubt it... The on-site trainer itself show previews of the training during the process and all of them vaguely look like the character in the dataset, wearing the same clothes and all, but its not the same identity at all, so I'm guessing the issue is that the default suggested training settings haven't been thought out yet. Or maybe does MinimaxH3 require a very specific set of training settings for lora character training ? Anyone ?

    dupe_900Aug 28, 2026
    CivitAI

    Can someone tell me what turbo fleet the CivitAi generator uses? The 600-step one?

    Checkpoint
    MiniMax H3
    by CivitaiOfficial
    Download (Beta) View on CivitAI
    base model

    Details

    Downloads
    8,746
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/10/2026
    Updated
    9/22/2026
    Deleted
    -

    Files

    minimaxH3_h3Comfy_nvfp4.safetensors

    Size:
    14.61 GB
    SHA256:
    35a88d51044231fe332301d7a62aa81e3f2cba62febeb446e2c1e3e0ef76f2c6

    Mirrors

    HuggingFace (20 mirrors)
    CivitAI (3 mirrors)

    minimaxH3_h3Comfy_3085597.safetensors

    Size:
    25.28 GB
    SHA256:
    bc2ced0fbea64757fa9acddccfc0b3f4819d1dcf1da6c124d690d368be283923

    Mirrors

    HuggingFace (8 mirrors)
    CivitAI (1 mirrors)

    minimaxH3_h3Comfy_3085684.safetensors

    Size:
    31.70 GB
    SHA256:
    7ad4c73e6e378b822ffd1629f27f632d3787d95f5e468e3af958f98c58df96a5

    Mirrors

    HuggingFace (3 mirrors)
    CivitAI (1 mirrors)

    minimaxH3_h3Comfy_3098344.safetensors

    Size:
    1.82 GB
    SHA256:
    fc9b6500f0331fe925b004738baaa31bd34104741c8bf9334816f9ac3005c8c1

    Mirrors

    HuggingFace (9 mirrors)
    CivitAI (2 mirrors)

    minimaxH3_h3Comfy_3098338.safetensors

    Size:
    300.29 MB
    SHA256:
    3a069f26fbc33f377a60dc72dd9e15f2aa42aa1d1b44915fded835716672dd36

    Mirrors

    HuggingFace (6 mirrors)
    CivitAI (4 mirrors)

    minimaxH3_h3Comfy_3086408.safetensors

    Size:
    19.53 GB
    SHA256:
    9255f52b6677845ad238f20dfaafa94727053694127ab7f255c048f0f9365779

    Mirrors

    HuggingFace (17 mirrors)
    CivitAI (2 mirrors)
    TensorArt (1 mirrors)
    TensorHub (1 mirrors)

    minimax_h3_audio_vae_fp32.safetensors

    Size:
    577.22 MB
    SHA256:
    8e505d95dd1561d47abd43d4238fd40d9bb1ae9e147ed0a4cba778d76ae4db48

    Mirrors

    HuggingFace (38 mirrors)
    CivitAI (2 mirrors)

    minimaxH3_h3Comfy_3098340.safetensors

    Size:
    300.29 MB
    SHA256:
    3a069f26fbc33f377a60dc72dd9e15f2aa42aa1d1b44915fded835716672dd36

    Mirrors

    HuggingFace (6 mirrors)
    CivitAI (4 mirrors)

    minimax_h3_video_vae_fp16.safetensors

    Size:
    4.85 GB
    SHA256:
    7c1f131492e7eddacaac9069a61b81bdd39de5cc96561e677c5eab1cdce5e522

    Mirrors

    HuggingFace (35 mirrors)
    CivitAI (3 mirrors)

    minimaxH3_h3Comfy_3085569.safetensors

    Size:
    47.97 GB
    SHA256:
    600d567f6a9629c8574e8e7041b199bdd9c59a986afa7906910a81919610607d

    Mirrors

    HuggingFace (3 mirrors)
    CivitAI (1 mirrors)

    minimaxH3_h3Comfy_3085784.safetensors

    Size:
    61.73 GB
    SHA256:

    Mirrors