Genmo open-sources 10B-parameter Mochi 1 video model
Genmo has released Mochi 1, a 10-billion-parameter text-to-video model under Apache 2.0 that it calls the largest openly released video model, and announced a $28.4 million Series A.
Genmo announced a research preview of Mochi 1, an open text-to-video diffusion model available under the Apache 2.0 license for both personal and commercial use. The weights and architecture are on Hugging Face, and a free hosted playground is live. The current release is a 480p base model. Genmo says a Mochi 1 HD version will follow later this year.
The model was trained from scratch on a new Asymmetric Diffusion Transformer (AsymmDiT) architecture. Its visual stream has nearly four times as many parameters as its text stream, a design Genmo says reduces inference memory. Prompts go through a single T5-XXL encoder. The model attends over a 44,520-token context with full 3D attention and learnable 3D RoPE. Genmo is also open-sourcing its video VAE, which compresses video 96x into a 12-channel latent space using 8x8 spatial and 6x temporal compression.
Mochi 1 generates 30 fps clips of up to 5.4 seconds. Genmo calls it best-in-class among open models and competitive with closed ones, but those claims come from the company's own evaluations. Prompt adherence was scored with Gemini-1.5-Pro-002 acting as a judge. Motion quality was measured with Elo ratings following the Chatbot Arena protocol. Genmo also lists known limitations: output is capped at 480p, extreme motion can cause warping, and animated styles come out poorly. A technical paper is promised.
NEA's Rick Yang led the $28.4 million Series A. The House Fund, WndrCo, and Replit CEO Amjad Masad were among the other investors.