Varya, Avataar's distilled video model designed for India

Avataar distills Alibaba's Wan 2.2 into Varya, a low-cost video AI model backed by the IndiaAI Mission to offer culturally accurate generation for India.

Backed by the IndiaAI Mission, a public program funded with approximately $1.2 billion, Indian developer Avataar is showcasing Varya, a video generation model built around two priorities: a very low cost and a strong understanding of the Indian context.

Rather than starting from scratch, the team distilled Wan 2.2, Alibaba's open-weights video model (14 billion parameters), reducing the number of denoising steps from fifty to four. On an H200 GPU, Varya then generates a five-second 720p clip in about forty seconds, whereas the original model requires more than twelve hundred.

Regarding pricing, the company claims 0.48 rupees (approximately $0.005) per second on its hosted service, which it presents as up to twenty times cheaper than Veo, Kling, Luma, or Runway. The other argument is cultural accuracy, encompassing Indian festivals, clothing, dishes, or architecture, whereas global models often lean towards the generic.

The quality, stated to be comparable to Wan 2.2 and Veo 3, remains for now an in-house claim: a technical report and the open weights are announced on the India AI Kosh portal, while the demonstration, in text-to-video and image-to-video, is already open.