What is MiniMax H3?
“Generate and edit videos up to 15 seconds long from as many as 15 references—across video, image, and audio. H3 follows complex instructions, transfers motion, and renders text beautifully. And it produces dialogue, music, and sound effects in native stereo.”
Source: Pika’s MiniMax H3 page. Icon: Pika.
MiniMax launched H3 on July 31, 2026, as “a general-purpose multimodal generation model.” It says H3 “understands unified context across text, images, video, and audio, generating video with native stereo sound, up to 15 seconds at 2K resolution.” On August 3, 2026, MiniMax open-sourced the model weights under its MiniMax H3 Community License Agreement.
Sources: MiniMax, MiniMax H3 announcement and open-source release.
Key features
Up to 15 seconds, up to 2K
Pika lists 15 seconds; MiniMax adds “at 2K resolution.”
Native stereo sound
“Dialogue, music, and sound effects in native stereo,” generated with the picture.
Up to 15 references
A mix of video, image and audio references in one generation.
Text and brand rendering
MiniMax highlights “accurate text and brand rendering”; Pika says it “renders text beautifully.”
Motion transfer
MiniMax lists “V2V motion transfer”: movement taken from a reference video.
Generate and edit
Pika says H3 can both generate and edit videos.
Pika’s page doesn’t list prices, aspect ratios or duration options for H3; the app shows them. See Pika pricing.
A MiniMax H3 workflow
- Open MiniMax H3 and sign in through Pika’s login page.
- Write the brief: what happens, in order; any spoken lines in quotes; the sound you want.
- Attach references: characters or products (images), motion (video), music or voice (audio).
- Choose length and shape from the options the app shows.
- Check the credit cost, then generate.
- Watch with sound on: check instructions, text, lip sync and stereo placement.
- Edit the clip to fix one thing rather than regenerating everything.
Prompt tips for MiniMax H3
- Write actions in order: “rises, loops twice, then dives.” H3 is built to follow complex instructions.
- Put dialogue in quotation marks and say who speaks each line.
- Keep on-screen text short and in quotes: “NIGHT OWL COFFEE.”
- Describe sound with the action: “the door slams,” “rain on the glass.”
- Name each reference’s job: “use @dance for the moves, @character for the look.”
- Add a “keep” line for faces, products and labels.
For general video prompting, see the text-to-video guide.
Using references
| Reference | Use it for |
|---|---|
| Image | Characters, products, locations, style |
| Video | Motion to transfer, or a clip to edit |
| Audio | Music, voice or sound design to follow |
Pika lists up to 15 references across video, image and audio; check the app for any limit per type. Use only material you have rights to, and people only with consent. For motion from a recorded performance, also see Motion Transfer.
Example prompts and results
Six prompts built around H3’s strengths: dialogue, text, motion transfer, complex instructions, audio references and editing. Turn your sound on. The prompts are our own.
Café dialogue scene
Two friends at a small café table by a rainy window. The woman leans in and says, “You actually quit?” The man smiles and replies, “Yesterday. Best decision I’ve made.” Medium shot, warm light, café ambience, rain on the glass.
What to look for: Lip sync on both lines and café sounds placed naturally in stereo.
Try one change: Change the second line and check the sync again.
Brand text on a sign
A neon sign reading “NIGHT OWL COFFEE” flickers on above a shop door on a wet street at night. Camera slowly pushes in. Buzzing neon sound, distant traffic.
What to look for: Every letter of the sign is correct and stays readable as the camera moves.
Try one change: Change the sign text to your own brand name, kept short.
Dance from a reference video
@character performs the moves from @dance in an empty rooftop at sunset. Match the timing of the reference exactly. Keep @character’s face and outfit as in the reference. Upbeat music.
What to look for: The moves follow the reference video’s timing while the character stays consistent.
Try one change: Swap the rooftop for a neon-lit studio.
Complex instruction
A red kite rises from a beach, loops twice to the left, then dives toward the sea before pulling up. A child runs along the shore following it. Wide shot, wind and waves, bright afternoon.
What to look for: Two loops to the left, a dive and a pull-up, in that order.
Try one change: Add one more instruction and see whether the order still holds.
Product with music
@bottle on a marble counter as sunlight moves across it, a slow orbit around it. Use @track as the music. Keep @bottle’s label and colour exactly as in the reference.
What to look for: The label is sharp and the music follows the reference track’s mood.
Try one change: Replace the audio reference with a calmer track.
Edit an existing clip
In @clip, change the daytime sky to a sunset and add warm light on the buildings. Keep the people, motion and camera exactly the same.
What to look for: Only the sky and lighting change; the motion is untouched.
Try one change: Try a different edit: turn the scene into light snowfall.
Known limits
MiniMax says “H3 still has room to grow” and lists priorities for future versions. Check fine detail, especially faces and hands, before publishing.
- 15 seconds is the limit per generation; use Extend Video or several clips for more.
- Long on-screen text is harder than a short sign or headline.
- Always check lip sync on every line of dialogue.
When to choose MiniMax H3
Choose H3 for
- Short scenes with dialogue and stereo sound
- Readable signs, logos and brand text in video
- Following detailed, step-by-step instructions
- Transferring motion from a reference video
- Editing a clip you already have
Consider another model for
- Up to 30-second multi-shot stories: Seedance 2.5
- Cheap drafts: Wan 3.0 (33 credits per 5s clip, per Pika pricing)
- Keyframe transitions: Pika 2.5
Compare all models in the video models guide.
MiniMax H3 FAQs
How long can MiniMax H3 videos be?
Up to 15 seconds per generation, according to Pika and MiniMax.
Does MiniMax H3 make sound?
Yes: “dialogue, music, and sound effects in native stereo.”
Can it put text in videos?
Yes. Pika says it “renders text beautifully” and MiniMax highlights “accurate text and brand rendering.” Keep text short and check it.
Is MiniMax H3 open source?
MiniMax released the model weights on August 3, 2026, under its MiniMax H3 Community License Agreement. In Pika, you use it through the app.
Is it the same as MiniMax Music 3.0?
No. H3 is MiniMax’s video model; MiniMax Music 3.0 is its music model.
Is this the official Pika or MiniMax website?
No. pikaais.com is an independent informational guide, not affiliated with Pika or MiniMax.
Make a short scene with dialogue
Copy prompt 01, generate it, and check the lip sync and café sound. Then change one line and compare.
Open MiniMax H3 ↗Next: Seedance 2.5 · Video models · AI video hub
Sources: Pika MiniMax H3, MiniMax’s H3 announcement and open-source release, and Pika pricing, checked September 24, 2026. Images and icon: Pika. Prompts and tips are our own. Not affiliated with MiniMax.