- Published on
An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio
- Authors
- Name
- DeepHorus
MiniMax has unveiled MiniMax H3, a general-purpose multimodal generation model that produces up to 15-second 2K video clips with native stereo audio. Released on July 31, 2026, the model is available through the MiniMax platform API under the model ID MiniMax-H3 and in the consumer Hailuo AI app. Unlike typical text-to-video systems with add-on audio, H3 reads text, images, video, and audio as one unified context and returns video with synchronized sound, marking a shift toward truly omni-modal content generation.
Historically, video generation stacks were split into separate expert models for text-to-video, image-to-video, first-and-last-frame, subject reference, motion reference, and video editing. MiniMax H3 folds these into a single pretraining paradigm where reference and editing relationships are expressed in natural language. For example, a user can prompt the model to reference camera movement from one video, have a character from an image sing, and match vocals to a separate audio clip—all in one request.
The API supports three entry modes: text-to-video, first/last-frame image-to-video, and reference generation, all behind one asynchronous endpoint. Input limits include up to 9 reference images, 3 reference videos (2–15 seconds each, totaling 15 seconds), and 3 reference audio clips, with a mixed input cap of 12 files. Prompt length is limited to 7,000 characters, and request bodies must stay under 64 MB; per-asset limits are 50 MB for video, 30 MB for images, and 15 MB for audio.
Four technical components power H3's capabilities. The Contextual Omni Representation rebuilds captioning to describe relationships between context and target video, distilling roughly 100K tokens of inference to about 4K tokens on average. The H3-VAE tokenizer delivers a 4× gain in effective sequence length, enabling native 2K output. The H3-Omni Transformer separates understanding and generation workloads, boosting end-to-end training throughput by nearly 30%. Finally, In-Context Regeneration allows the base model to regenerate its own low-resolution output while re-reading the original multimodal context, recovering small text and fine detail that conventional upscalers miss.
MiniMax positions H3 for advertising, branding, e-commerce, product design, UI/UX, gaming, film pre-visualization, and retail catalog media. Applications include ad variant generation, product videos, animated posters, film title sequences, website hero loops, and character-consistent game cinematics.
The company claims that at 2K, H3's per-second price is less than a third of mainstream models, and at 768p, less than half the price of mainstream 720p. Third-party trackers report a 2K pay-as-you-go rate of $0.13 per second, or about $1.95 for a 15-second clip, though MiniMax's official pricing page still listed only Hailuo 2.3 tiers at the time of writing. According to SCMP, citing Artificial Analysis, H3 leads in video editing benchmarks, indicating strong performance against established rivals.
While H3 is deployable today via API, on-premises deployment is not yet available, potentially limiting adoption among enterprises with strict data governance requirements. MiniMax says the model is live now, and the multi-format support for H.264/H.265 video, JPG/PNG/WEBP/HEIC/HEIF images, and WAV/MP3 audio suggests broad compatibility. As the platform matures, H3 could redefine how creators generate video and audio together, but pricing transparency and deployment flexibility will be key to wider industry acceptance.
