Via tomsguide.com
The Chinese AI company's open-weight video model is the first of its kind to top an Artificial Analysis video editing leaderboard, generating 2K clips with synchronized stereo sound.
A 33-billion-parameter model from Chinese AI firm MiniMax has taken the top spot in the Video Edit Arena on arena.ai, scoring 1390 points and opening a 32-point gap over its closest competitors, Dreamina Seedance 2.0 and Gemini Omni Flash. The model, called MiniMax-H3, launched on July 31, 2026, and needed roughly two weeks to climb to the number one position.
It’s also the first open-weight model to dominate a video editing category on the Artificial Analysis leaderboards. The weights are hosted on Hugging Face, though access is restricted for US individuals and certain other users.
What MiniMax-H3 actually does #
MiniMax-H3 is what the industry calls an “omni-modal” video generation model. You feed it text, images, videos, audio, or some combination of all four, and it produces 2K-resolution video clips ranging from 4 to 15 seconds with synchronized stereo sound baked in.
A single prompt can reference up to 9 images, 3 videos, and 3 audio files simultaneously. That level of multi-input flexibility is a big part of why the model excels at instruction-based editing tasks like subject replacement and relighting.
The architecture underneath is a diffusion transformer. At 33 billion parameters, it can run through an API priced at $7.80 per minute of generated video. MiniMax hosts the consumer-facing version inside its Hailuo AI app.
The leaderboard picture #
Video Edit Arena isn’t the only benchmark where MiniMax-H3 is making noise. According to Artificial Analysis leaderboards, the model currently holds the top position in video editing overall, places second in text-to-video generation, and ranks third in image-to-video tasks.
The 32-point margin over Dreamina Seedance 2.0 and Gemini Omni Flash in video editing is substantial by arena standards, where models often cluster within single-digit differences.
Why open weights matter here #
Most of the models competing at the top of AI video leaderboards are closed-source, meaning users interact through an API or app but never see the model’s internals. MiniMax-H3 breaks that pattern by releasing open weights on Hugging Face, at least for users outside restricted jurisdictions.
Open weights allow researchers and developers to fine-tune the model for specific use cases, run it on their own infrastructure, and inspect how it handles different inputs. For commercial applications in advertising, e-commerce, and UI/UX design, where companies often need to customize AI tools for proprietary workflows, this is a meaningful advantage over black-box alternatives.
The restriction on US access limits the model’s reach in the world’s largest advertising market and creates a two-tier ecosystem where some developers can build on MiniMax-H3 directly while others are confined to the API.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our