Avataar AI cuts video generation costs by 20x with new Varya model

By distilling Alibaba’s Wan 2.2 model, Avataar transformed a process previously requiring 50 steps into a lean four-step pipeline. This engineering shift allows an NVIDIA H200 GPU to render a 5-second 720p clip in 45 seconds—a massive leap from the 1,230 seconds required by the original architecture. This efficiency enables the startup to price its services at ₹0.48 per second, targeting the specific needs of Indian enterprises, educators, and small business owners who have been priced out of Western alternatives.

Beyond raw speed, the model addresses a recurring flaw in generative AI: cultural misalignment. Avataar curated its training data to ensure Varya accurately depicts Indian festivals, regional attire, and local architecture, moving away from the generic outputs typical of global models. The startup plans to release Varya as an open-weight model on India’s AI Kosh portal, inviting developers to self-host or integrate the technology into their own platforms. This move aligns with a broader national strategy to prioritize application-level innovation and developer ecosystems over the costly, compute-heavy race to build foundational models from scratch.