Fine-Tuning Friday: Exploring MiniMax H3 w/ Fillip Isgro (Machine Delusions)

​Welcome to this week’s edition of our favorite series from the Oxen.ai Herd, Fine-Tune Fridays.

​We share practical, end-to-end examples, including reference data, model weights, and the full infrastructure needed to reproduce the experiments on your own.

​This Week

​Running Seedance 2.0 locally is no longer a dream! (Or that's what people are saying at least). Let's see how true that is by putting MiniMaxAI/MiniMax-H3 to the test with Fill!

​MiniMax-H3 is a 33B parameter open-weights video model that takes the same kinds of inputs as Seedance, text, images, video, and audio references, all in a single prompt. It handles text-to-video, image-to-video, and editing in one model, generates clips up to 15 seconds with synchronized audio, and folks are already getting it running on a single consumer GPU.

​Looking forward to seeing you there!