Mochi is a video generation model that produces high-fidelity video from text prompts using an open-source diffusion architecture.
The tool addresses the gap between closed commercial video generation systems and open alternatives by implementing a 10 billion parameter Asymmetric Diffusion Transformer model trained from scratch. It pairs this with an AsymmVAE encoder-decoder that compresses video to a 128x smaller latent representation, enabling efficient inference. The model is released under Apache 2.0 and can run on consumer GPUs, with native integration into ComfyUI for accessible deployment.
Developers should choose this tool if they need an open-source video generation model they can run locally or fine-tune on their own data. It suits projects requiring strong prompt adherence and high-fidelity motion without reliance on proprietary APIs. The repository provides multiple interfaces: a Gradio UI for interactive generation, a command-line tool for batch processing, and a Python API for programmatic use. Fine-tuning via LoRA is supported, allowing adaptation to custom video styles on a single H100 or A100 GPU. The straightforward, hackable architecture makes it suitable for researchers and practitioners wanting to understand or modify the underlying model.
The project maintains active development with recent additions of LoRA fine-tuning support and consumer GPU optimization. The codebase is well-documented with example scripts for both API usage and fine-tuning workflows. Installation is streamlined through standard Python package management, and model weights are distributed through multiple channels including Hugging Face and direct download. The repository includes an efficient context parallel inference implementation alongside the core model, indicating attention to practical deployment concerns.
