Open-Weight Models by License

Eighteen licenses cover everything in this hub: 2,760 models, 859 datasets and 254 papers from 1,692 publishers. That ratio is the point. A team that wants to run open weights in its own facility has to clear the license before it clears the hardware, and reading 1,692 publishers one at a time is a job nobody finishes. Sorted by license, the catalog becomes 18 questions. We read the license line as carefully as the memory figure because it decides what the weights are for once they sit on our disks: who may use what they produce, whether a modified copy can leave the building, and what the publisher keeps a say over.

Work the page from the license inward. Pick the terms your counsel will sign, open that license, and only then sort by task: text generation is the largest at 338 models, with image and text to text at 138 and robotics at 81. When license and task line up, the model page carries the memory figure and the hourly GPU price we read on Sep 18, 2026. Costing out a model you cannot deploy is the mistake this order prevents.