Open-Sora

Open-Sora is HPC-AI Tech's open source effort to democratize video generation. Its 2.0 release is an 11 billion parameter model that handles text-to-video and image-to-video in a single network, producing clips of up to 129 frames, roughly five seconds, at 256px or 768px in aspect ratios including 16:9, 9:16, 1:1, and 2.39:1. The team reports quality on par with models such as HunyuanVideo on VBench and human preference evaluations, and stresses cost efficiency, stating the model was trained for about 200,000 dollars. The repository publishes weights, training and inference code, and data processing pipelines, so the project functions as a reproducible reference for video diffusion research rather than only a demo. Inference requires a GPU with substantial VRAM, and the earlier 1.x releases remain available for smaller setups. Researchers and independent builders use it to study, fine-tune, and extend open video generation.