MiniMax released H3 on July 31 — an open-weight, omni-modal generation model that produces native 2K video at 24FPS, in clips from 5 to 15 seconds, with native stereo audio (dialogue, sound effects, and music) generated in the same pass as the picture rather than added afterward by a separate audio model.
The resolution trick worth understanding
Most models that ship at higher resolutions get there with a dedicated super-resolution module bolted onto a lower-resolution base generation — a separate pass that upscales, often losing fine detail like small text or intricate textures in the process. H3 does something different: for its 2K output, the base model regenerates its own low-resolution output in-context, effectively re-rendering the scene at full detail rather than stretching an existing low-res frame. MiniMax’s own framing is that this recovers detail — small text, fine texture — that a conventional super-resolution step typically can’t restore, because it’s generating that detail directly rather than inferring it from an already-degraded source.
What you can actually feed it
H3 supports three entry modes: straightforward text-to-video, first/last-frame image-to-video (specifying both the starting and ending frame and letting the model generate the motion between them), and reference generation — up to 9 reference images, 3 reference videos, and 3 reference audio clips can steer a single generation toward a specific look, motion style, or sound. On top of generation, H3 supports instruction-based editing: describing a change to an existing image, video, or audio clip in natural language rather than re-generating from scratch or manually compositing a fix.
Why “omni-modal” is the accurate description, not marketing language
MiniMax’s framing is that H3 “understands unified context across text, images, video, and audio” — and the reference-generation feature is the clearest evidence that’s a real architectural property rather than a tagline. Steering a single video generation with reference images and reference videos and reference audio simultaneously requires the model to actually reason across those modalities jointly, not just accept each as an isolated conditioning input bolted onto an otherwise text-driven pipeline.
Pricing and access
MiniMax is pricing H3 at under a third the cost of mainstream competitors for comparable 2K output with audio, with the full weights set to open-source — meaning, unlike a closed API-only release, researchers and toolmakers will be able to build directly on the model rather than only calling it as a service. For a site tracking both commercial AI creative tools and the open-source ecosystem around them, an open-weighted model matching or undercutting closed competitors on price and capability is the more consequential half of this release.
Related Reading
- MiniMax Releases MiniMax H3: An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio — MarkTechPost
- MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities — MiniMax (official)
- MiniMax H3 - Open-Weights General-Purpose Multimodal Video Model — fal
- China’s MiniMax Open-Sources H3 Video Model — Value Add VC