MiniMax Releases MiniMax H3: Omni-Modal Video Model Generating 15-Second 2K Clips with Native Stereo Audio

MiniMax released MiniMax H3, a general-purpose multimodal generation model that reads text, images, video, and audio as a unified context and returns video with native stereo audio, supporting 2K resolution, durations from 4 to 15 seconds (integers only). Unlike previous stacks that required separate models for text-to-video, image-to-video, and other tasks, H3 folds these into a single pretraining paradigm where relationships are expressed in natural language, such as referencing camera movement from a video or matching vocals to audio. The model launched on July 31, 2026, available via API under the model ID MiniMax-H3 and in the consumer Hailuo AI app. It is positioned for advertising, branding, e-commerce, product design, UI/UX, gaming, film pre-visualization, and retail catalog media, with applications including ad variants, product videos, animated posters, and video-to-video motion transfer. The API supports three entry modes—text-to-video, first/last-frame image-to-video, and reference generation—with input limits including up to 9 reference images, 3 reference videos (2-15 s each, ≤15 s total), and 3 reference audio clips (audio requires accompanying image/video). Mixed input caps at 12 files; prompts ≤7,000 characters; request body ≤64 MB; file sizes: video ≤50 MB, image ≤30 MB, audio ≤15 MB. Formats include H.264/H.265 video, JPG/PNG/WEBP/HEIC/HEIF images, and WAV/MP3 audio. Four technical components enable its performance: Contextual Omni Representation (captioning that describes relationships, distilling ~100K tokens to ~4K), H3-VAE (a tokenizer with 4× effective sequence-length gain enabling native 2K), H3-Omni Transformer (separates understanding and generation workloads, boosting training throughput ~30%), and In-Context Regeneration (base model regenerates low-res output in-context instead of bolting on super-resolution). Pricing claims: at 2K, per-second price is under a third of mainstream models; at 768p, under half of mainstream 720p. Third-party trackers report $0.13 per second (~$1.95 per 15-second clip), but MiniMax's pay-as-you-go page still lists Hailuo 2.3 tiers. According to SCMP citing Artificial Analysis, H3 leads in video editing, trails Google's Gemini Omni Flash in text-to-video, and sits behind Seedance 2.0 and Gemini Omni Flash in image-to-video. Open weights are promised 'in the coming days' but not yet shipped.