MiniMax H3 is an open-source omni-modal generation model from Shanghai AI company MiniMax, built for video creators, ad producers and engineering teams that want commercial-grade footage without paying for a closed API. Announced on 31 July 2026 and open-sourced on 3 August, H3 reads text, images, audio and video inside one multimodal context and returns up to 2K output, 15 seconds long, with native stereo audio generated in the same pass.
Core Features
- Unified omni-modal input: mix text prompts, reference stills, audio clips and source video in a single request
- Direct 2K output up to 15 seconds with native stereo sound, so no separate audio stage is needed
- V2V Motion Transfer copies motion from a reference clip while swapping the actor, wardrobe, product or set
- Ranked first for video editing on the Artificial Analysis video model leaderboard
- Three-part system (H3-Context-IR, H3-Base, H3-Regenerate-2K) downloadable from Hugging Face and ModelScope
- Runs on SGLang, vLLM, diffusers and ComfyUI, with day-one silicon support from Huawei Ascend, AMD, Intel, Moore Threads and Biren
Use Cases
- Ad and e-commerce teams generating product spots that keep brand text and logos legible
- Studios producing character or wardrobe variants through motion transfer instead of a reshoot
- Enterprises deploying locally to keep footage and training data inside their own network
- Game and UI/UX teams prototyping animated concepts before committing art budget
Pricing
Weights are free to download from Hugging Face and ModelScope, so self-hosting costs only compute. Hosted 2K generation runs about RMB 0.8 per second (roughly USD 0.11), which MiniMax states is a third of comparable flagship video models. Cloud inference is also offered through fal and RunningHub.
Our Take
H3 is the most credible open-weight answer to Sora and Veo so far. The 15-second ceiling is limiting for narrative work, but motion transfer plus native audio at a third of the going rate makes it the sensible pick for teams shipping high volumes of commercial edits. Compare it with other releases at aifreetool.site/tool-category/ai-video/
