Mochi

Mochi is Genmo's open-source text-to-video generation model designed to convert written descriptions into video sequences. Unlike image-to-video models that start from static frames, Mochi accepts text prompts and generates entire videos from scratch based on the semantic content of the prompt. The model is listed here specifically because of its open weights and source code, which enable local, self-hosted deployment rather than reliance on a commercial API. While Genmo operates a web interface at genmo.ai for interactive testing and playground features, the underlying model weights and inference code are available publicly, making Mochi a genuine open-source option within the text-to-video landscape. This approach reflects Genmo's commitment to participating in the open AI research community while maintaining a commercial service offering.

The open weights live on Hugging Face at huggingface.co/genmo/mochi-1-preview, and the complete source code is available on GitHub at github.com/genmoai/mochi. This dual availability means researchers and developers can access both the trained model and the implementation details needed for deployment. The Apache 2.0 license applied to Mochi provides permissive terms for both academic and commercial use, subject to the standard conditions of that license. Users can modify the code, build upon it, and integrate it into proprietary systems without complex licensing negotiations. The approach reflects Genmo's decision to participate in the open-source model ecosystem rather than restricting access to a closed service. The decision to open-source a capable video generation model demonstrates confidence in the technology and builds goodwill within the research community.

Mochi generates video at 480p resolution with a 10-billion-parameter diffusion transformer architecture. The model uses an Asymmetric Diffusion Transformer (AsymmDiT) that separates text and visual processing streams, allowing different dimensionality for text embeddings (1,536 dimensions) and visual tokens (3,072 dimensions). This asymmetry reflects the observation that visual information requires higher dimensional representation than text in the transformer architecture. The video compression system, implemented through an AsymmVAE, achieves 128x compression of the input space, using 8x spatial and 6x temporal reduction to produce a 12-channel latent representation. This architecture balances generation quality with computational efficiency, allowing reasonable inference times on high-end GPUs. The 10-billion-parameter count places Mochi in the mid-size range for modern diffusion models, making it larger than smaller efficient variants but smaller than some research-scale implementations.

Running Mochi locally demands significant GPU memory. Single-GPU operation requires approximately 60GB VRAM, with an H100 listed as the recommended hardware. However, optimizations through ComfyUI integration can reduce requirements to under 20GB through techniques like attention optimization and sequential processing. Diffusers implementations with BF16 precision can operate in approximately 22GB, albeit with slight quality reduction compared to full precision. The software requirements include Python with UV package manager and FFMPEG for video processing and encoding. Installation follows standard open-source practices, with documentation guiding users through setup and dependency management. The availability of multiple deployment paths, from high-end A100 clusters to optimized consumer GPU setups, makes Mochi accessible to organizations with different resource constraints.

The model supports multiple usage methods tailored to different workflows. A Gradio-based interactive interface provides a web UI for experimentation with features like negative prompt conditioning and CPU offload for systems with limited VRAM. Negative prompts allow users to steer generation away from unwanted elements, providing finer control over the output. A command-line interface enables batch processing and integration into scripts and pipelines, supporting automation and production workflows. A Python API offers programmatic access through composable pipeline components, allowing developers to build custom workflows and integrate video generation into larger applications. The project also includes LoRA fine-tuning support, enabling users to adapt the model for specific domains or visual styles. LoRA (Low-Rank Adaptation) fine-tuning requires substantial GPU resources like H100 or A100 80GB but allows domain-specific customization without full model retraining.

Mochi demonstrates particular strengths with photorealistic content and prompt adherence. The model generates motion-rich videos that generally maintain consistency with the text prompt throughout the sequence. The AsymmDiT architecture's focus on motion quality means generated videos prioritize smooth, believable motion over perfect photorealism. Variable frame counts are supported during inference, enabling users to trade video length for computational cost. Users can request different lengths and frame rates, making the model adaptable to various content requirements. However, the model carries documented limitations. It does not perform well with animated content, struggling when prompts call for stylized or non-photorealistic art, cartoon characters, or illustrative styles. Extreme motion scenarios can introduce minor warping or distortions. Like all generative models, Mochi reflects biases present in its training data, primarily photorealistic content, and Genmo recommends that organizations implement additional safety protocols when deploying in production contexts.

Genmo's playground at genmo.ai/play offers a risk-free way to explore Mochi's capabilities before committing to local deployment. The web interface handles all computation on Genmo's infrastructure, removing the need to provision GPUs for initial experimentation. This approach serves as both a marketing tool and a genuine utility for users evaluating the model. The playground shows video generation in motion and allows rapid iteration on prompts, helping users understand what the model can and cannot produce. Internally, the playground uses the same model weights available open-source, ensuring fidelity between the public service and the downloadable version. The availability of a free playground reduces the barrier to entry for artists and creators considering whether to adopt Mochi for their projects.

Community adoption of Mochi has driven integration into popular creative workflows. ComfyUI support allows visual node-based construction of video generation pipelines. Integration with tools like Runway and other video editing platforms is ongoing, expanding accessibility. The project's GitHub repository maintains active development with regular updates and community contributions. The Apache 2.0 license enables researchers to fork the project and explore variations, and several research groups have already extended Mochi's architecture for specific applications. This ecosystem of tools and integrations extends the model's usefulness beyond basic inference.

Mochi positions itself within the open-source text-to-video space as a model prioritizing motion quality and prompt understanding over photorealism extremes or ultra-high resolution. The 480p resolution represents a practical choice balancing generation quality and speed, though it means output requires upscaling for publication in higher resolutions. The project is deliberately placed in the open-source-video-models category rather than in a separate commercial text-to-video category because the emphasis is on the open, self-hostable model itself, not on Genmo as a hosted service. For teams and researchers seeking a capable, locally-deployable text-to-video generator with permissive licensing and active development, Mochi represents a mature option. The combination of open weights, clear documentation, multiple usage modes, and reasonable performance makes Mochi well-suited for production use where openness and customization matter as much as output quality. Genmo's continued investment in the open model while running a commercial service demonstrates a balanced approach to building sustainable business models around open-source AI.