MiniMax releases omni-modal generation model H3: 2K video at 1/3 the price of mainstream models
赛博禅心 · wechat · 2026-07-31
MiniMax has officially released H3, a general-purpose omni-modal generation model that understands and generates text, images, video, and sound, with native dual-channel audio and video output up to 15 seconds at 2K resolution. H3 excels in instruction following, text and brand presentation, and video-to-video motion transfer, making it suitable for advertising, e-commerce, gaming, and more.
Technical highlights:
- Contextual Omni Representation: Enhances multi-modal context understanding, unifying task descriptions via language.
- H3-VAE: High compression yields 4x sequence length efficiency, reducing training and inference costs.
- H3-OmniTransformer: Heterogeneous architecture for understanding and generation, boosting end-to-end training throughput by 30%.
- In-context Regeneration: Uses the base model to regenerate low-res outputs, surpassing traditional super-resolution.
Pricing: Under 1/3 the per-second cost of mainstream models at 2K, and under 1/2 at 768P.
Open-source plan: Weights to be released in the coming days, with compatibility for domestic chips.
Design philosophy: Breaking task boundaries, moving from specialized to general, unifying image, video, and audio generation tasks.
Related event: MiniMax Releases Multimodal Model H3 with 2K Native Stereo Video(18 posts)→
More from Multimodal
- Fan animation made in Blender with Claude Code (Opus 5.5) — Icy-Appointment3583 · 2026-10-16
- Making Blender animations with Opus, Astra and Argon: how you prompt still matters — chongdashu · 2026-10-16
- Tessellation-free geometric queries on displaced surfaces cut caustics variance 10x — ssh4net · 2026-10-16
- Complex multi-link CAD mechanism finally animates, dev celebrates the breakthrough — zetalyrae · 2026-10-16
- OpenMOSS releases MOSS-Image-SFT: Apache 2.0 text-to-image model on Hugging Face — bdsqlsz · 2026-10-16
- Hyperflow open-sourced: 49 sampling steps down to 8, ~3x faster video generation — mhdfaran · 2026-10-16