MiniMax H3 in Genosai — video with sound from text, frames and references
MiniMax H3, also known as Hailuo 03, is MiniMax's video model that generates a clip together with stereo sound: ambient atmosphere and effects. It works in three modes — text to video, first and last frame animation, and reference-based generation — produces clips from 4 to 15 seconds in 768P or 2K, and supports six aspect ratios. It has been available in Genosai since September 1, 2026, online with no subscription — from 7 credits per second.
Updated: September 27, 2026
- Sound generated with the picture — A stereo track with ambience and sound effects is created in the same pass as the video — no need to add audio separately.
- Three modes in one model — Text to video, first and last frame animation and reference-based generation — the studio picks the mode automatically from your uploads.
- Up to 9 images, 3 videos and 3 audio files — In reference mode the model accepts images, video clips up to 15 seconds in total and audio tracks for the character, style and sound.
- From 4 to 15 seconds — Any whole-second duration in that range, with 768P for drafts and 2K for the final clip.
- Six aspect ratios — 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9, plus an adaptive ratio that follows your sources in reference mode.
Contents
- What is MiniMax H3
- Capabilities
- Examples and use cases
- How to use it in Genosai
- Prompts
- Generation cost
- Comparison
- Limitations and tips
- FAQ
What is MiniMax H3
MiniMax H3 is a video model from the Chinese company MiniMax, also known as Hailuo 03. It is the next generation after the Hailuo 2.3 line: the previous model, Hailuo 2.3, runs in Genosai only in image-to-video mode, while H3 can start from text, from a frame or from a set of references — and adds sound to the picture right away.
The headline feature of this generation is a native stereo track. The model generates ambient atmosphere and sound effects in the same pass as the video: rain taps on the umbrella exactly when the drops fall in frame, and ice clinks at the moment it touches the glass. That saves a post-production step: no hunting for library sounds and no nudging them into sync on a timeline.
The second difference is input flexibility. One model combines three modes: text to video, animation of a first and/or last frame, and reference-based generation, where you can pass up to 9 images, up to 3 video clips and up to 3 audio tracks at once. Duration is set to the second — from 4 to 15 — and resolution is 768P or 2K.
MiniMax H3 has been available in Genosai since September 1, 2026. The model runs online: no developer account, no foreign service keys and no VPN. You open the video studio in the browser, describe the scene — and get a clip with sound.
Capabilities
MiniMax H3 covers the typical short-video cycle: from an idea in text or a ready frame to a clip with sound in the format you need. Here is what the model does in practice.
Sound is generated together with the video
Every clip comes with stereo sound: the atmosphere of the place, footsteps, rain, street noise, clinking dishes, fizzing soda. The sound is synced with events in the frame because the model creates it at the same time as the image rather than layering it on top. The more specifically you describe the soundscape in the prompt, the more precise the track: "soft rain on a transparent umbrella and a distant city hum" works better than just "with sound".
Three generation modes
Text to video is the classic mode: you describe the scene in words, and the model invents the frame, the motion and the sound. Image to video lets you upload a first frame, a last frame or both, and the model builds the motion between them; this is handy for animating product photos and making seamless transitions. Reference mode lets you hand the model a set of materials, and it assembles a new scene from them. In the Genosai video studio the mode is picked automatically based on what you upload, so there is nothing to switch manually.
References as images, video and sound
In reference mode MiniMax H3 accepts up to 9 images, up to 3 video clips with a combined length of up to 15 seconds and up to 3 audio tracks. Images define the character's look, the product and the style; video defines the nature of the motion and the camera work; audio defines the sound atmosphere. An audio reference only works together with an image or video: the model does not build a clip from sound alone. The aspect ratio in this mode can be left adaptive so the model follows your sources.
Duration, resolution and aspect ratios
Duration can be any whole number from 4 to 15 seconds, so you do not overpay for extra seconds just to fit a fixed grid. Resolution is 768P for quick idea checks and 2K for the final clip. Aspect ratios are 16:9 for YouTube and presentations, 9:16 for Reels, Shorts and stories, 1:1 and 4:3 or 3:4 for social feeds and marketplaces, and 21:9 for a cinematic wide frame.
Detailed prompts
A prompt can be up to 7,000 characters long. That is enough to describe the character, setting, lighting, camera movement, second-by-second action and soundscape without trimming the description to fit a limit.
Examples and use cases
Product clips for marketplaces and social media are one of the most cost-effective jobs for MiniMax H3. A vertical clip with a product where you hear ice clinking, bubbles fizzing or a cap clicking feels more alive than a silent one, and it needs no sound designer. If you already have a product photo, upload it as the first frame — the model keeps the packaging and adds camera motion.
Atmospheric scenes and intros benefit from sound even more: a rainy night city, a forest at dawn, a cafe kitchen at rush hour. For branded content, reference mode works well: a character photo, the product and a camera-motion example come together in one scene. First and last frame mode helps with transitions between two states — before and after, a change of season, an unboxing.
How to use it in Genosai
Running MiniMax H3 in Genosai requires no setup: everything happens in the browser.
- Sign in at genosai.io and open the video studio.
- Pick MiniMax H3 in the model list.
- Describe the scene: character, action, camera movement, lighting and sounds.
- Optionally upload a first and last frame or references — the mode is selected automatically.
- Set a duration from 4 to 15 seconds, 768P or 2K resolution and the aspect ratio.
- Check the cost on the button, start the generation and download the finished clip.
Through the API
For automation the model is available in the Genosai public API. Create a task with a POST request to https://api.genosai.io/v1/createTask using the model minimax-h3, then fetch the status and the video link with GET /v1/taskInfo?taskId=. One or two images in image_urls switch on frame mode, while three or more images or the reference_image_urls field switch on reference mode.
{
"model": "minimax-h3",
"input": {
"prompt": "A glass bottle of lemonade on a wooden bar counter, sound of ice clinking",
"aspect_ratio": "9:16",
"resolution": "768P",
"duration": 6
}
}Prompts
A few rules noticeably improve the result. Start with the subject and the location, then describe the action and camera movement in concrete terms — slow dolly, orbit, push-in, static wide shot. Describe the sound in a separate phrase at the end: which sources, how loud, what sits in the foreground. Set the lighting and time of day — "neon on wet asphalt", "bright summer daylight". Below are real examples: the prompt itself and the clip MiniMax H3 generated from it in Genosai (768P, 6 seconds, 42 credits per clip).
A cinematic street scene, 16:9, 6 seconds, text to video:
Cinematic slow dolly shot along a rainy night street in a modern city, neon reflections on wet asphalt, a woman in a yellow raincoat walks away from camera holding a transparent umbrella, light rain, ambient city sounds and soft rain on the umbrella.Rainy night street, 16:9, 6 seconds — generated in Genosai.io with MiniMax H3
The model delivered a 1344×768 clip: neon reflections on wet asphalt, a heroine in a yellow raincoat who stays consistent through the whole walk, and the sound of rain over the city background.
A vertical product video, 9:16, 6 seconds, text to video:
Vertical product video: a glass bottle of lemonade with ice cubes on a wooden bar counter, a hand places a fresh lemon slice on the rim, condensation drops slide down the glass, the camera slowly orbits around the bottle, bright summer daylight, sound of ice clinking and fizzing bubbles.Lemonade product video, 9:16, 6 seconds — generated in Genosai.io with MiniMax H3
The result is a 768×1344 vertical: a hand places a lemon slice on the rim of the bottle, and the track carries the sound of clinking ice. A clip like this can go straight into a product card or Reels.
Generation cost
MiniMax H3 is billed per second in Genosai, and the rate depends on resolution: 7 credits per second at 768P and 12 credits per second at 2K. Sound is included in the price — there is no separate charge for it.
| Duration | 768P | 2K |
|---|---|---|
| 4 seconds | 28 credits | 48 credits |
| 6 seconds | 42 credits | 72 credits |
| 10 seconds | 70 credits | 120 credits |
| 15 seconds | 105 credits | 180 credits |
The cost-effective approach is to refine the idea, composition and sound on short 768P clips and run only the final version in 2K. The exact cost is shown on the button before each generation. You pay per actual generation with no subscription; top-ups accept Russian bank cards, and starter credits after sign-up let you try the model for free. Current packages are listed on the Pricing page.
Comparison
Within the MiniMax family: Hailuo 2.3 in Genosai works only from a ready frame, while MiniMax H3 starts from text, frames or references, adds a stereo track and stretches the clip to 15 seconds.
| Model | Strength | Price in Genosai |
|---|---|---|
| MiniMax H3 | Sound, three modes, 4–15 s, 2K | 7 credits/s at 768P, 12 at 2K |
| Hailuo 2.3 | Frame animation up to 1080p | from 45 credits per clip |
| Kling 3.0 | Multi-shot editing up to 15 s | from 14 credits/s |
| Veo 3.1 Fast | 4–8 s clips with sound, up to 4K | from 60 credits per clip |
| Wan 3.0 | Up to 30 s in one pass, speech | from 8 credits/s |
| Seedance 2.0 | References and sound up to 1080p | from 9 credits/s |
If you need a single long scene of up to 30 seconds, look at Wan 3.0 or Seedance 2.5. If multi-shot directing within one generation matters, try Kling 3.0. For a short clip of up to 8 seconds with a 4K option, Veo 3.1 Fast fits. MiniMax H3 pays off where you need a predictable short clip with sound at a clear per-second price and a flexible choice of duration. All models are available in one interface, so it is easy to compare them on your own scene.
Limitations and tips
Duration is capped at 15 seconds per generation — build longer videos from several scenes. An audio reference only works together with an image or video. When animating frames, the uploaded image sets the aspect ratio of the clip, so prepare the source in the format you need. Through the API, reference mode accepts images and audio, while video references currently work only in the studio.
Practical tips: start with 768P to test the idea cheaply, and switch to 2K for the final. Always describe the sound in the prompt — without a hint the model picks the atmosphere on its own. For product clips, use a real product photo as the first frame so the packaging matches the original. And always review the result before publishing: a neural network can distort small text on packaging, the number of fingers or a logo, and it is better to catch such details before your customers see the clip.
FAQ
What is MiniMax H3 and how is it related to Hailuo?
MiniMax H3, also known as Hailuo 03, is a video model from MiniMax and the successor to the Hailuo 2.3 line. Its headline differences are native stereo sound generated together with the picture and three working modes: text, frames and references. It has been available in Genosai since September 1, 2026.
How much does a MiniMax H3 video cost in Genosai?
7 credits per second at 768P and 12 credits per second at 2K. A 6-second 768P clip costs 42 credits, a 10-second one costs 70, six seconds at 2K cost 72, and the maximum 15 seconds at 2K cost 180 credits. You pay per generation, with no subscription.
Which generation modes does MiniMax H3 support?
Three modes: text to video, image to video from a first and/or last frame, and reference-based generation with up to 9 images, up to 3 videos totaling up to 15 seconds and up to 3 audio tracks. In the Genosai video studio the mode is chosen automatically based on what you upload.
What durations and resolutions are available?
Duration ranges from 4 to 15 seconds, any whole number. Resolution is 768P or 2K. Aspect ratios are 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9, plus adaptive in reference mode. The prompt can be up to 7,000 characters long.
What kind of sound does MiniMax H3 generate?
The model creates a stereo track together with the video: ambient atmosphere and sound effects such as rain, clinking ice or city hum. Describe the sounds you want right in the prompt. An audio reference can only be added together with an image or video reference.
Can I use MiniMax H3 through an API?
Yes. In the Genosai public API the model is called via POST /v1/createTask with the model minimax-h3, and the status is checked with GET /v1/taskInfo. One or two images switch on frame mode, while three or more images or the reference_image_urls field switch on reference mode.
Do I need a subscription?
No. Genosai charges per actual generation with no monthly subscription, and top-ups accept Russian bank cards with no VPN required. Starter credits after sign-up let you try the model for free.