Space Bunny Alpha was an early MiniMax M3.1 Flash. This playground now runs the official model — launch credit packs are live.

Home/Updated 2026-11-03

MiniMax H3

MiniMax H3 is MiniMax’s open, omni-modal video model. It takes text, images, video, and audio in one context and generates video with native stereo audio — up to about 15 seconds, with vendor 2K paths on the hosted stack. spacebunny.me is an independent guide hub: we do not generate H3 video on this server.

Not affiliated with MiniMax, ComfyUI, fal.ai, or Hugging Face. Specs below cite the Hugging Face model card, MiniMax API docs, and Comfy Org’s day-0 post.

What it is, in three sentences

  1. Open-weight omni-modal system: unified multimodal context in, video with native stereo audio out (Hugging Face MiniMax-H3 model card).
  2. Clip length is listed as 4–15 seconds at 24 FPS; short edge defaults to 768 pixels, and vendor 2K generation uses H3-Regenerate-2K, which is not in the open H3-Base release (MiniMax local-deploy docs).
  3. You can try hosted UIs, download weights, call fal/MiniMax APIs, or run ComfyUI 0.30+ templates that ship quantized Comfy-Org files.

Demo videos

Embedded public demos. The first is from Comfy Org; the others are community walkthroughs that show real local output. MiniMax’s own sample gallery lives on their open-source news page.

MiniMax H3 day-0 support in ComfyUI — Comfy Org on YouTube. Official Comfy Org livestream covering open weights, native stereo audio, and the T2V / I2V / R2V templates.
Local MiniMax H3 overview and install walkthrough — Community tutorial on YouTube. Community walkthrough of text, image, and reference workflows with native audio. Not a MiniMax-produced clip — useful for seeing local output.
How to use MiniMax H3 in ComfyUI — Community tutorial on YouTube. Step-by-step ComfyUI usage covering prompts, first/last frame, and reference packs.

Official write-up: MiniMax H3 is now open source.

Specs and limits

ItemValueSource / note
Duration4–15 s (H3); 5–15 s (H3 Max API)MiniMax API / model card
Frame rate24 FPSModel card
Audio32 kHz native stereo with the videoModel card
Open H3-Base resolutionShort edge 768 px default (~1344×768 at 16:9)HF card + MiniMax local-deploy guide
Vendor 2K pathH3-Regenerate-2KNot included in the open H3-Base release
Open weightsYes — FL2VA and Ref2VAMiniMax H3 Community License
vs Hailuo 3 / Hailuo AISame generation often sold as Hailuo 3 in the product UIHailuo product copy; Hailuo 2.3 remains a legacy API line
vs H3 Maxfal post-train of H3; 480p/768p, faster hosted pathMiniMax API models intro + fal comparison

Prompt starters

Five copy-ready prompts. Full set and a browser generator: MiniMax H3 prompts.

  • Editorial tech product film

    Reference images + text · Adapted from Comfy Org day-0 examples

    Editorial tech product film. The transparent gaming mouse from <Picture 1> in its original scene: a pitch-black studio void with a dark, subtle reflective surface, lit by dramatic duotone vibrant blue and warm neon orange rim lighting.
    SHOT 1: open exactly on image 1; slow push-in as the lights pulse through the acrylic shell.
    SHOT 2: extreme macro of the scroll wheel; warm orange beam sweeps the metal internals.
    SHOT 3: low-angle beauty shot; the mouse levitates a few centimeters and rotates slowly, then fades to silhouette.
    Audio: deep pulsing sub-bass, sharp tactile clicks, glassy whoosh on cuts.
  • Bold comic-book action with synced SFX

    Reference frames + audio · Adapted from Comfy Org day-0 examples

    Bold comic-book ink style, heavy linework, red and blue-black palette, night city. Use <Picture 2> and <Picture 1> as reference frames and <Audio 1> exactly as it is.
    CUT 1: top-down view of the little boy superhero on the rooftop — red cape fluttering, cocky grin into camera — comic overlay text word by word: "GET READY TO" / "MEET" / "YOUR" / "MAKER".
    TRANSITION: violent whip pan that smears the floating words.
    CUT 2: low hero angle on a colossal black mech-kaiju roaring over the skyline; speed-lines and ink splatter at the roar peak.
  • High-fashion editorial with multi-image refs

    Three reference images · Adapted from Comfy Org day-0 examples

    High-fashion editorial film, luxurious slow motion, soft gradient studio sky.
    MUSIC & SFX: cinematic score with deep taiko drums, shimmering koto plucks, modern sub-bass.
    SHOT 1: beside her, the mask hangs broken into the floating shard formation of <Picture 2>.
    SHOT 2: gold seams ignite; shards snap together until the mask fuses.
    SHOT 3: the golden dragon of <Picture 3> swoops through the frame.
    SHOT 4: the mask pulls onto her face; glowing kintsugi veins spread down the jacket.
    SHOT 5: she lands on the dark liquid wave in a warrior stance — hold as a lookbook frame.
    Use <Picture 1>, <Picture 2>, <Picture 3> as reference images.
  • Fisheye product commercial

    Single reference image · Adapted from Comfy Org day-0 examples

    Vibrant fisheye product commercial, hyper-saturated summer light. The woman from <Picture 1> in a yellow raincoat crouched by a jungle waterfall holding a rainbow-gradient soda can toward the lens.
    MUSIC: upbeat tropical house — punchy kick, steel-drum plucks, warm bass.
    CUT 1: fisheye hero frame; giant typography stamps "STAY" then "HYDRATED" behind her on the beat.
    TRANSITION: extreme close-up of the tab — click-hiss — iris shutters to black.
    CUT 2: iris reopens; she dumps the liquid, rainbow refraction through the stream.
    CUT 3: rainbow can floating in the turquoise pool; "STAY COMFY" as a water reflection. Hold product hero frame.
  • Text-only rainy alley (no refs)

    Text-to-video · Spacebunny original starter prompt

    Photoreal night alley in Osaka, light rain, neon reflections on wet asphalt. A courier in a yellow jacket parks a bicycle, checks a phone, then looks up as a train passes overhead. Handheld camera, shallow depth of field, 8 seconds, continuous shot, no cuts. Native stereo: distant train rumble, rain on metal, muffled city murmur. No logos, no readable text on signs.

Where you can run it

  • MiniMax / Hailuo product — Hosted product UI. Creators often call the same model Hailuo 3 or H3.
  • Hugging Face weights — Official open checkpoints: FL2VA and Ref2VA under the MiniMax H3 Community License.
  • fal.ai hosted API — Hosted H3 and H3 Max endpoints with published per-second prices. Spacebunny does not proxy this API.
  • ComfyUI — Day-0 templates for T2V, I2V, and R2V. Uses Comfy-Org quantized files, not the raw BF16 HF dump.

Local install detail: MiniMax H3 ComfyUI guide.

Hosted cost estimate

Pure frontend math from fal’s published per-second rates. Local ComfyUI has no API fee.

Per-second cost calculator

Uses fal.ai published USD/second rates for hosted H3 / H3 Max. Local ComfyUI is $0 in API fees but still costs electricity and VRAM. Not a MiniMax invoice.

≈ $0.48($0.06/s × 8s)

Questions

Is MiniMax H3 any good?

For open weights, it is one of the strongest omni-modal video options published so far: native stereo audio, multimodal references, and day-0 ComfyUI support. Quality still depends on the path — hosted 2K API vs local H3-Base at a 768-pixel short edge are not the same product. Spacebunny has not published an independent leaderboard score.

What is MiniMax H3 used for?

Text-to-video, image-to-video, first/last-frame control, and reference-to-video that can carry subject, motion, or voice from images, clips, and audio. Typical uses: product ads, comic-style shorts, fashion editorial cuts, and local ComfyUI experimentation.

How long can MiniMax H3 generate?

MiniMax’s model card and API docs list 4–15 seconds for MiniMax H3 (integer seconds). H3 Max on the MiniMax API is listed as 5–15 seconds. Clips run at 24 FPS.

How censored is MiniMax H3?

Hosted MiniMax / Hailuo and fal endpoints apply their own safety filters; what is blocked depends on the product, not only the weights. Local open weights follow the MiniMax H3 Community License and whatever filters you add yourself. This site does not claim a complete censored/uncensored matrix.

Related MiniMax H3 pages

spacebunny.me is independent and not affiliated with MiniMax, ComfyUI, fal.ai, or Hugging Face. This hub does not generate video on our servers.