genmoai/mochi is a repository for a video generation model that creates videos matching a given sentence. Under the Apache-2.0 license, the weights of Mochi 1, the text-conditioned video generation model, the VAE, and the inference implementation were all released together. It is written in Python and has 3752 stars. No released version has been confirmed, and the repository was created on September 11, 2024.
The README explains that Mochi 1 demonstrated motion fidelity and prompt adherence. Text-conditioned generation is the method in which the input sentence determines the content of the video, and prompt adherence refers to how well that sentence is reflected in the video. The model uses the AsymmDiT architecture with 10 billion parameters, and AsymmDiT stands for Diffusion Transformer structure. Video generation is offered through three paths — Gradio UI, CLI, and API — and the repository includes a VAE (Variational Autoencoder, an autoencoder that converts data into a compressed representation and restores it) and the inference implementation. It also supports LoRA (Low-Rank Adaptation, a method that adds new traits to trained weights with only a few parameters) fine-tuning using your own video data.
However, the resources needed to run it are substantial. According to the README, running it on a single GPU requires about 60GB of VRAM. The motion fidelity and prompt adherence figures are the preliminary evaluation results cited by the README. This record contains no independently measured numbers, no performance benchmarks, and no comparison figures against other video models. The recorded momentum index is 2.5 out of 10, and the freshness, technology, and quality scores are empty. It is a subject worth checking directly on hardware with room to spare, and any judgment beyond that needs more supporting evidence.