What is mochi?

Mochi 1 is an open-source video generation model created by Genmo. You give it a text description and it produces a short video clip. It is positioned as one of the most capable freely available video generation models, closing the distance between what open-source projects can do and what proprietary commercial systems offer. The model is very large, with 10 billion parameters, and is built on an architecture Genmo calls AsymmDiT (Asymmetric Diffusion Transformer). The basic idea is that a diffusion model starts from noise and gradually refines it into a final output. This model handles text descriptions and video frames together in one process, spending more of its processing capacity on the visual side of the problem. It compresses video into a much smaller internal representation first (using a separate component called AsymmVAE) and then works on that compressed form before expanding it back out into actual video frames. Videos are currently generated at 480p resolution. Running it requires substantial hardware. On a single graphics card you need roughly 60 GB of video memory, which puts it out of reach for most consumer GPUs. The recommended setup is at least one H100, a high-end data center GPU. The README notes that the ComfyUI integration (a popular visual interface for AI image and video tools) can bring that requirement down to under 20 GB, though that path is separate from this repository. To use it locally, you download the model weights separately (they are available on Hugging Face or via a direct link), then run either a browser-based graphical interface built with Gradio or a simpler command-line script. A Python API is also included if you want to call the model from your own code. The project added support for LoRA fine-tuning in late 2024, which means you can adapt the model to a specific visual style by training it further on your own video examples, though that step still requires a high-end GPU such as an H100 or A100. The model is released under the Apache 2.0 license, which is permissive for both personal and commercial use. Genmo notes that the model reflects biases present in its training data and recommends additional safety review before any commercial deployment.