MiniMax H3 Review: Full‑Modal Video Editing Beats After Effects for Just 0.5 CNY per Second

MiniMax H3, the newly released flagship multimodal video model, has quickly become a hot topic after its open‑source launch, with early testers reporting SOTA results on the Artificial Analysis leaderboard and claiming the top global spot for video editing with audio.

The model’s API pricing is strikingly low—text‑to‑video generation costs only 0.5 CNY per second, and even 2K resolution video costs just 0.8 CNY per second, clearly undercutting competitors such as Seedance 2.0.

Key capabilities include full‑modal input, precise multimodal editing and control, output up to 2K resolution and 15 seconds length, and native dual‑channel audio generated together with the visual content. Text generation quality is markedly improved over the previous generation, reliably producing complex text such as subtitles, titles, logos and promotional copy, while the model’s voice‑cloning ability also receives praise.

To evaluate real‑world performance, the authors used the classic "华强买瓜" clip. First, a 300‑word text description was fed to MiniMax H3, producing a video that preserved the original characters’ expressions and story tension. Adding two original character screenshots as an additional image modality allowed the model to recreate the scene with the same cast.

Further tests replaced all characters with cats using a single textual command, changed the aspect ratio, and generated an English‑dubbed version, demonstrating that both visual and audio elements can be edited simultaneously. The model also created a new happy ending for the clip, maintaining accurate lip‑sync, natural facial expressions, and consistent audio quality.

Additional experiments showed multi‑photo outfit swaps, virtual travel scenes, and the ability to turn a storyboard image directly into a short film.

These results illustrate MiniMax H3’s "AI‑native post‑production" capability: unlike the fragmented past workflow—separate text‑to‑video, image‑to‑video, face‑swap, background‑replace, motion‑transfer, and separate audio pipelines followed by manual compositing in tools like After Effects—MiniMax H3 handles generation, editing, layout, and dubbing within a single model.

The underlying technical foundations are:

Contextual Omni Representation : describes relationships among all input modalities and the target video.

H3‑VAE tokenizer : a redesigned tokenizer that offers higher compression, enabling native 2K output.

H3‑Omni Transformer : a task‑generalized architecture that separates understanding and generation workloads, improving training efficiency.

In‑context Regeneration : the model re‑examines its low‑resolution output together with the original context to regenerate high‑resolution details.

These components collectively enhance the model’s ability to understand context, which is essential for the "AI native post‑production" vision where generation and editing are unified.

The open‑source nature of MiniMax H3 is expected to accelerate ecosystem adoption: developers can adapt the model to various Chinese chips, private‑deploy it behind firewalls to protect proprietary assets, and avoid the data‑leak risk of closed‑source APIs. Open‑source flagship models also encourage faster iteration and broader community contributions compared with the slower, closed‑source video‑generation landscape.

MiniMax H3 is positioned as the first AI video model that directly challenges After Effects by offering end‑to‑end generation and post‑production capabilities, marking a significant shift toward an open, collaborative video‑generation ecosystem.

Prompt example: style: dark‑pop / cyber‑grunge / rap music video, realistic high‑fashion texture, film‑magazine feel, high contrast but not cheap, referencing late‑90s to early‑00s indie magazines, with grain, slight film shake, halftone dots, printed edges, scan misalignment; fast cuts only, no fades; text packaging style and texture as in reference image.

Overall, MiniMax H3 demonstrates that a single open‑source model can generate, edit, and post‑process video content with quality and control previously achievable only with a suite of specialized, closed‑source tools.