Advertising & e-commerce
Hero product orbits, UGC-style demos, pack shots with readable logos. 2K holds up on a product page; 9:16 ships to Reels the same afternoon.
MiniMax H3 Video · 2K · Native stereo
MiniMax’s omni-modal video model, live on Nano Banana. Text or a photo in. A 4–15s clip out, with native stereo sound, brand-accurate type, and cinematic motion.
No card · 25 credits · 768P or 2K · Commercial rights
“An influencer presents a beauty product”
MiniMax H3 Video specs, as shipped on Nano Banana
2K
or 768P
4–15s
24 FPS
32 kHz
native stereo
6 ratios
21:9 → 9:16
33B
H3-Omni Transformer
11
dialogue languages
What is MiniMax H3 Video?
MiniMax H3 Video (Hailuo 3) is a general-purpose generative system from MiniMax. It was launched in July 2026 and open-weighted in August 2026. One model reads text, images, video and audio, then writes picture and stereo sound together.
Earlier Hailuo models split text-to-video, image-to-video, motion transfer and audio into separate tasks. H3 was trained to collapse those boundaries. That is why it follows complex briefs, renders brand type, and transfers camera motion from a reference without a pipeline of specialist models.
On Nano Banana you run MiniMax H3 Video in the browser: Studio, Banana Computer, or the API. No GPUs, no Hugging Face weights, no MiniMax Open Platform key.
H3 was built around Contextual Omni Representation: language is the bridge between modalities. Multi-shot briefs, camera notes and brand constraints stay intact instead of being averaged away.
Picture and 32 kHz stereo are generated together. Voice, effects and music are one audio stream, not a silent clip you score later.
H3-Base writes 768P. H3-Regenerate-2K feeds that result plus the original context back into the same model, recovering small type and product detail that a super-res upscaler would invent.
Early commercial tests highlight accurate on-screen text and logos. Product sites, title cards and UI mockups are first-class MiniMax H3 Video jobs, not an afterthought.
H3-VAE compresses video and audio with a 4× gain in sequence length, which is what makes native 2K tractable. The H3-Omni Transformer is a 33B dense single-stream model. In-context regeneration, not a bolted-on upscaler, produces the 2K pass.
Same composer as Banana Computer. Select MiniMax H3, describe the shot, attach a still if you have one. The agent can search, draft stills, then render MiniMax H3 Video in the same thread.
video_model: "veo3"
No MiniMax console, no 4-GPU local setup, no Community License geography questions. Pick the model and generate.
In Studio or Banana Computer, select MiniMax H3. That is the MiniMax H3 Video engine. API id veo3.
Text-to-video from a sentence. Image-to-video from a product shot, frame or poster. Duration 4–15s, ratio including 21:9 and 9:16.
Download an MP4 with native stereo. Paid plans include commercial rights. Ready for ads, product pages and social.
Official MiniMax H3 output, mapped to what you can actually select in Nano Banana Studio.
| Model | MiniMax H3 Video (Hailuo 3) · API veo3 |
|---|---|
| Duration | 4–15 seconds (Studio: 5 / 8 / 10 / 12 / 15) |
| Resolution | 768P (H3-Base) or 2K (H3-Regenerate-2K) |
| Frame rate | 24 FPS |
| Audio | Native 32 kHz stereo, generated with the picture |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Inputs on Nano Banana | Text-to-video and image-to-video |
| Architecture | 33B H3-Omni Transformer · H3-VAE · in-context 2K regen |
| Dialogue languages | Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish |
MiniMax designed H3 for commercial work: ads, product, brand type, motion transfer. These are the jobs we see it used for on Nano Banana.
Hero product orbits, UGC-style demos, pack shots with readable logos. 2K holds up on a product page; 9:16 ships to Reels the same afternoon.
Opening cards, animated posters, 21:9 cinematic stills that move. H3’s in-context 2K pass is the reason to pick it when on-screen type has to stay readable.
Dusk exteriors, interior walkthroughs, material close-ups. 2K plus stereo room tone reads as a real location scout, not a slideshow.
App walkthroughs and website heroes where on-screen UI must stay sharp. In-context 2K is built to keep small interface type intact.
Lookbook motion, texture, steam, rain, handheld or locked-off. Image-to-video from a still campaign frame is a one-click MiniMax H3 Video job.
Short narrative beats with synced stereo. Pair H3 for the 2K hero, Seedance 2.5 for a 30-second take, Sora 2 when dialogue and physics have to lock, same workspace.
Official model specs. Use MiniMax H3 Video when you need 2K and in-context regeneration for small type.
| Dimension |
MiniMax H3 Video
MiniMax · Hailuo 3
|
Seedance 2.5
ByteDance
|
Sora 2
OpenAI
|
|---|---|---|---|
| Best for | 2K heroes, brand type, ads, architecture | Long takes, product, multi-reference, realistic people | Physics, lip-synced dialogue, narrative |
| Max resolution | 2K (768P native) | 480p / 720p | 720p · 1080p on Pro |
| Duration | 4–15 s | 4–30 s | Up to 20 s |
| Audio | Native 32 kHz stereo | Native, joint with picture | Native synced (dialogue, SFX, ambience) |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | 16:9 and 9:16 |
| References | Text + image (on Nano Banana) | Up to 50 (images, video, audio) | Text + image |
| API id | veo3 |
seedance2 |
sora2 |
Straight answers for people searching MiniMax H3 Video.
video_model: "veo3" on text-to-video or image-to-video requests. That id is MiniMax H3. Docs: AI models.
MiniMax H3 Video · Live on Nano Banana
2K, native stereo, 4–15 seconds. Type a prompt in Banana Computer or lock settings in Studio.