Open Weights, Top Ranking
Chinese AI company MiniMax has released the open weights of its H3 video model, making it the first open model to top an AI video ranking. The release came on Aug 3, 2026, the same day ByteDance launched its closed competitor, Seedance 2.5. The ranking comes from Artificial Analysis, which now places H3 first in Video Editing, second in Text-to-Video, and third in Image-to-Video.
H3 is a 33-billion-parameter model. It processes text, images, video, and audio together in a single architecture. That unified approach lets it generate clips of four to 15 seconds with stereo sound. A single prompt can include up to nine reference images, three video clips, and three audio clips, giving users a wide canvas for complex scenes.
What the Open Release Includes
The model weights are hosted on HuggingFace, and the open release covers the core generation pipeline. Users can run H3 locally through ComfyUI, a tool for running AI models on personal hardware. However, local use tops out at 768p resolution, a clear step down from the model's full capabilities.
Two pieces of H3 remain closed. The 2K resolution module is not included in the open release, and neither is H3-Context-IR, which translates prompts and reference material into a structured intermediate format. That means users will need to handle context prep themselves using MiniMax's published prompting guides.
Fine-Tuning and Licensing
The open weights allow fine-tuning on custom footage, characters, or a specific visual style. That flexibility is a major draw for studios and independent creators who want a model tailored to their own material. But the license carries a revenue cap. Commercial use is only permitted for companies making under $20 million in revenue, a threshold that keeps the model accessible to smaller teams while limiting larger enterprises.
That $20 million limit could shape adoption. Startups and indie filmmakers can experiment freely, while bigger players must weigh the cost of licensing or building their own tools. The restriction is a common pattern in open-weight releases, but it still draws attention in a market where commercial video generation is heating up.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
ByteDance's Same-Day Answer
ByteDance did not wait long to respond. On the same day, it released Seedance 2.5, a closed model that generates 30-second clips with built-in audio. That is double the maximum length of H3's four-to-15-second output, giving ByteDance a clear edge in long-form generation. Seedance 2.5 is closed, so users cannot fine-tune it or run it locally, but its length advantage may appeal to filmmakers who need sustained scenes.
The timing is notable. Two major Chinese AI firms shipped competing video models on the same date, one open and one closed. That split reflects a broader strategic divide in the industry, where some companies bet on community-driven adoption and others protect their technology behind APIs.
The State of Open Video Models
H3's top ranking in Video Editing is a milestone. No open model had previously led an AI video benchmark, and the achievement signals that open weights can match or beat closed systems in at least some tasks. The second-place finish in Text-to-Video and third in Image-to-Video reinforce that H3 is not a one-trick model.
Still, the gaps remain. The missing 2K module and H3-Context-IR mean local users get a reduced experience, and the 768p ceiling in ComfyUI limits output quality. For those who want the full resolution and automated context handling, the closed version remains out of reach.
The article, written by Maximilian Schreiner for The Decoder, includes a video example by MiniMax H3. That example shows the model's ability to blend multiple inputs into a coherent clip with synchronized audio. The source material for the report came from HuggingFace, where the weights are publicly available.