MiniMax has launched MiniMax H3, a multimodal generation model that produces 2K video clips between four and 15 seconds long with native stereo audio built in. The system treats text, images, video, and sound as a single context rather than separate inputs. It outputs only integer durations.
How it works
Earlier video tools relied on separate models for text-to-video, image-to-video, frame matching, and editing. MiniMax H3 combines these functions into one pretraining approach where instructions use natural language. A prompt might ask the model to copy camera movement from one clip, have a character from a second image sing, and match the vocals to a third audio file.
Availability and pricing
Deployment is currently possible via the API. Users cannot run the model on their own hardware. The launch occurred on July 31, 2026. The model ID is MiniMax-H3. It is available through the platform API and the Hailuo AI consumer application.
Target industries include advertising, branding, e-commerce, product design, UI/UX, gaming, film pre-visualisation, and retail catalog media. Specific applications cover ad variant creation, product listing videos, animated posters, film title sequences, website hero loops, consistent character cinematics for games, and video-to-video motion transfer.
Technical specifications
The video generation guide outlines three entry modes: text-to-video, first-and-last-frame image-to-video, and reference generation. A single endpoint handles all requests using an asynchronous three-step flow: create a task, poll the task ID, and download the content URL.
Input limits for developers include:
- Up to nine reference images.
- Up to three reference video clips, each between 2 and 15 seconds, with a total duration limit of 15 seconds.
- Up to three audio clips. Audio files must accompany an image or video.
- A maximum of 12 input files in total.
- Prompt length must not exceed 7,000 characters.
- Request body size must stay under 64 MB. Large assets should use URL input.
- Individual file size limits are 50 MB for video, 30 MB for images, and 15 MB for audio.
- Supported formats include H.264 and H.265 for video, JPG, PNG, WEBP, HEIC, and HEIF for images, and WAV and MP3 for audio.
Core architecture
MiniMax rebuilt captioning to describe the relationship between context and the target video. This process reduces inference tokens from roughly 100K to about 4K on average. Language acts as the bridge to turn fixed tasks into open descriptions.
The H3-VAE is a full tokenizer overhaul. Its high compression ratio provides a fourfold gain in effective sequence length. This reduction cuts training and inference costs and enables native 2K output.
The H3-Omni Transformer separates understanding and generation workloads. The training architecture adjusts hardware usage for each task. This change boosted end-to-end training throughput by nearly 30 percent. The system explicitly moved away from the Hailuo-02 architecture.
In-Context Regeneration replaces bolt-on super-resolution modules. The base model regenerates its own low-resolution output by re-reading the original multimodal context. This approach preserves small text and fine detail that standard upscalers often miss, which matters for brand and product rendering.
Cost and performance
MiniMax claims the per-second price at 2K is less than a third of mainstream models. At 768p, the cost is less than half that of mainstream 720p models. The company promoted these figures on X.
Third-party trackers report the 2K pay-as-you-go rate at $0.13 per second. That totals approximately $1.95 for a 15-second clip. However, the official pricing page only listed Hailuo 2.3 tiers at the time of writing, so treat the $0.13 figure as a reported estimate rather than a confirmed primary rate.
Placement data from SCMP, citing Artificial Analysis, shows H3 leads in video editing. It trails Google’s Gemini Omni Flash for text-to-video tasks. It also sits behind both Seedance 2.0 and Gemini Omni Flash for image-to-video generation.
What it means
Creators now have a single endpoint for complex compositing tasks that previously required stitching together multiple tools. The native stereo audio removes the need for separate lip-sync or sound design steps. The lower cost per second makes high-resolution output accessible for short-form commercial work.
