MiniMax H3 is MiniMax's audio-video generation model, open-sourced on August 3, 2026: one prompt gives a 4–15 second clip with synced sound and dialogue (24 fps, 32 kHz stereo, 768p by default, 2K through MiniMax's hosted API). This free builder writes prompts in H3's own structured format (timed shots, (S1) says: <d>[English] …</d> dialogue, overall_soundscape, non_diegetic_music), gives you the matching API request JSON, and shows the list price for your clip. It doesn't generate video itself: paste the prompt into the Hailuo AI web app, MiniMax Design, ComfyUI or your own H3 setup. The structure matters most for local open-weight runs, because MiniMax's hosted prompt-rewriting step (H3-Context-IR) isn't included in the open release.
How to use this MiniMax H3 prompt builder
- Pick the mode. Text to video needs no image. First frame, first + last frame and last frame modes describe motion around your keyframe images, and the API sets the ratio to adaptive to match them.
- Write one row per shot. The first shot has no time stamp. Each later shot needs a cut time that is later than the one before and shorter than the clip. Add a speaker and a line if someone talks: the line is kept word for word inside
<d>[Language] …</d>, and each speaker keeps the same ID (S1, S2). - Describe sound and music separately. overall_soundscape is ambience and action sounds; non_diegetic_music is score only the audience hears (leave it blank for N/A).
- Copy the prompt into the Hailuo AI web app, MiniMax Design, ComfyUI or your local H3 setup, or copy the JSON for the MiniMax API.
Why the structured format matters for local H3
On MiniMax's own API and apps, a hosted step called H3-Context-IR rewrites your idea into this structured form before generation. That step is not in the open-weight release, and MiniMax says it is critical to output quality. If you run the open weights yourself (ComfyUI, diffusers, SGLang, vLLM), you have to write prompts in that structure yourself — which is what this builder does. Sources: MiniMax-H3 model card, MiniMax open-source announcement.
MiniMax H3 at a glance
| MiniMax H3 | |
|---|---|
| What it makes | Video with audio (speech, ambience, music) from text, keyframes or references |
| Clip length | 4–15 seconds, 24 fps, 32 kHz stereo audio |
| Resolution | 768p short side by default; 2K via MiniMax's hosted API (H3-Regenerate-2K) |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Dialogue | 11 languages MiniMax lists as stable |
| Open weights | Released August 3, 2026 on Hugging Face (MiniMaxAI/MiniMax-H3) and GitHub (MiniMax-AI) |
| Model size | 33B dense Omni-Transformer; Qwen3-VL-32B text/vision encoder |
Where you can use MiniMax H3
| Option | What you need | Cost |
|---|---|---|
| Hailuo AI web app | Browser, MiniMax account | Credits / plan in the app |
| MiniMax Design (desktop app from MiniMax Hub) | Windows or Mac | App plan |
| MiniMax API | API key, model MiniMax-H3 |
Pay per second (below) |
| Open weights: SGLang, vLLM, diffusers, ComfyUI | Large data-centre GPUs (the official SGLang example uses 4 GPUs); tens of GB download | No API fee, but see the licence |
MiniMax H3 API price (pay-as-you-go list price)
| Model / output | Price |
|---|---|
| MiniMax-H3, 768P | $0.08 per second |
| MiniMax-H3, 2K | $0.13 per second |
| MiniMax-H3-Max, 480P / 768P | $0.05 / $0.08 per second |
| Reference images | First 5 per task free, then $0.04 each |
| Audio references | Free |
| Regenerate 768P → 2K | $0.05 per second |
A 10-second 2K clip is therefore $1.30; the same clip at 768P is $0.80. The builder above does this maths for your settings.
Licence limits for the open weights
- MiniMax's licence for the open weights excludes the EU, the UK, South Korea and the USA from its "Applicable Territory". If you are there, use the hosted app or API instead of running the weights.
- Companies with more than $20 million in annual revenue need MiniMax's authorisation for commercial use.
- Commercial products built on it must show "MiniMax H3" in the user interface.
- ComfyUI's documentation says commercial use of outputs from local runs needs a MiniMax commercial licence (available through Comfy).
Read the LICENSE file in the official repository before you rely on it; this is a summary, not legal advice.
Frequently asked questions
What is MiniMax H3?
MiniMax's audio-video generation model. From text, keyframe images or reference clips it makes a 4–15 second video with synced dialogue, sound effects and music, at 24 fps with 32 kHz stereo audio.
Is MiniMax H3 open source?
Yes, the weights were released on August 3, 2026 on Hugging Face (MiniMaxAI/MiniMax-H3) and GitHub, under MiniMax's own licence. That licence excludes the EU, UK, South Korea and the USA, and large companies (over $20M revenue) need permission for commercial use.
Can I use MiniMax H3 online for free?
This builder is free and needs no sign-up, but it only writes prompts. Generating the video happens in MiniMax's Hailuo AI web app, MiniMax Design or the API, which use MiniMax's own credits or per-second pricing.
How much does the MiniMax H3 API cost?
$0.08 per second at 768P and $0.13 per second at 2K (pay-as-you-go list price). The first 5 reference images per task are free, then $0.04 each; audio references are free.
What hardware do I need to run MiniMax H3 locally?
Data-centre-class GPUs: it is a 33B model with a 32B encoder, and the official SGLang example uses 4 GPUs. The ComfyUI template also downloads tens of gigabytes. Most people are better off with the hosted app or API.
Why do local H3 results look worse than the official app?
MiniMax's hosted service first rewrites your idea with a prompt model called H3-Context-IR, which isn't in the open release, and MiniMax says it is critical to quality. Writing prompts in H3's structured shot/dialogue/soundscape format, as this builder does, closes part of that gap.
Does this tool upload my prompt anywhere?
No. Everything runs in your browser; nothing is sent to us or to MiniMax until you paste it into their app or API yourself.