Artificial intelligence has become one of the world’s biggest technology races and India is now catching up with the US, Europe and China. Avataar AI, one of the 12 startups selected under the India AI Mission, has launched Varya, a new video generation AI model that aims to make video creation faster, cheaper and more relevant to users in India. The model is trained to better understand India’s cultural background, from food and clothing to festivals and architecture, areas where global AI models sometimes struggle.
Built on global technology optimized for India
Varya is based on Wan 2.2, a publicly available video generation model released by Alibaba. Avatar AI says it has compressed the functionality of Wan 2.2 into a smaller, faster version designed for its unique requirements. This process, known as model distillation, allows developers to reduce computing requirements while preserving much of the original functionality.
According to the company, this optimization has significantly improved speed. Varya runs in 4 steps instead of 50 steps in Wan 2.2. This enables up to 10x faster video generation while significantly reducing costs.
To put this in practical terms, Avataar AI says that with the NVIDIA H200 GPU, Varya can produce a 5-second 720p video in about 45 seconds, compared to 1,230 seconds for Wan 2.2.
The company claims that in its internal benchmarks, Varya produces videos at around Rs 0.48 per second, making it significantly more cost-effective than some of the leading global video models.
The company hopes to use AI to promote understanding of India
For users, the bigger difference may not be speed. Avatar AI says it trained Varya using a curated dataset to help the model better recognize India’s cultural nuances. That includes understanding what festivals look like, identifying local clothing styles, recognizing food, and interpreting architectural details more naturally. The goal is to reduce the situations in which AI-generated content looks generic or lacks cultural detail.
How Varya works
The experience is designed to be simple. Users can enter text prompts or upload images. The model then generates a video clip, and the user can continue to extend the sequence by generating additional clips.
Avatar AI also said it plans to publish a technical report describing the model’s architecture, distillation methodology, and benchmark results.
India’s AI ambitions are growing
Varya also positions itself as an open ecosystem project. The company plans to release it as an open weight model through AI Kosh, the Indian government’s central repository of publicly available AI models and datasets. This means that developers can host the model themselves, modify it, and build applications on top of it.
– end
