Mochi 1 Preview is an open-source AI video generator that creates high-fidelity motion videos from text prompts.
- for
- Developers, researchers, and creators who need customizable video generation.
- pricing
- open source
- license
- Apache-2.0
Key features
- High-Fidelity Motion — Produces industry-leading motion quality with a 10B parameter diffusion model.
- Open Source Architecture — Released under Apache 2.0 with the Mochi 1 Asymmetric Diffusion Transformer.
- Advanced Compression — Uses a VAE to compress videos 128× smaller with 8×8 spatial and 6× temporal reduction.
- Efficient Processing — Runs on a single T5-XXL language model for streamlined text handling.
- Multimodal Architecture — Joint attention to text and visual tokens via dedicated MLP layers and non-square QKV projection.
- Developer-Friendly — Simple, hackable codebase with comprehensive docs and community support.
Use cases
- Generate cinematic video clips from textual descriptions.
- Research and experiment with state-of-the-art video diffusion models.
- Create visual assets for games, films, or marketing materials.
- Compress generated videos for storage or distribution.
Mochi 1 Preview - AI video Generator vs alternatives
 Mochi 1 Preview - AI video Generator |  Compress |  BuildShip | |
|---|---|---|---|
| Best for | Open-source high-quality AI video generation | Video size reduction | AI-driven UI creation |
| Pricing | Open source | Free | Free |
| DevHunt upvotes | 0 | 55 | 118 |
| Launched | Nov 2025 | Mar 2024 | Jan 2023 |
- Mochi 1 Preview - AI video Generator vs Compress: Focuses on compressing existing videos rather than generating them.
- Mochi 1 Preview - AI video Generator vs BuildShip: Provides a visual backend builder for AI apps, not video generation.
Mochi 1 Preview - AI video Generator FAQ
What makes Mochi 1 unique?+
It combines a 10B parameter diffusion model with an asymmetric transformer architecture, delivering high-fidelity motion and strong prompt adherence.
What are the technical specifications of Mochi 1?+
Mochi 1 uses a 10B parameter AsymmDiT diffusion model, a single T5-XXL language model, and an open-source VAE for 128× video compression.
How does the Mochi 1 architecture work?+
It jointly attends to text and visual tokens with separate MLP layers and non-square QKV projection, enabling multimodal generation.
What are the current limitations of Mochi 1?+
The documentation notes it is in preliminary evaluation, so prompt adherence and motion quality may still evolve.
How can I run Mochi 1?+
Clone the repository, install dependencies with uv, configure model parameters, then generate videos via the Gradio UI or CLI.
Summarized by DevHunt from mochi1preview.com · Sep 28, 2026. Details may change; check the official site.
