India has been making strides in artificial intelligence (AI), yet its model output has lagged behind that of the U.S., Europe, and China. To boost innovation, the Indian government initiated the India AI Mission, a $1.2 billion initiative aimed at supporting startups with subsidized GPU computing in exchange for publicly shared models. Among the twelve startups selected for this initiative is Avataar AI, which has unveiled its new video model, Varya, designed to grasp local contexts such as festivals, food, and traditional attire.
Backed by Peak XV, Avataar AI has not developed Varya from the ground up. Instead, it utilized Wan 2.2, a publicly available video generation model from Alibaba, and applied a technique known as distillation. This process compresses the model's capabilities into a more efficient version, tailored for Avataar's unique applications. Consequently, Varya operates in four steps compared to the 50 steps required by Wan 2.2, resulting in video production that is ten times faster and significantly more cost-effective.
For instance, with an NVIDIA H200 GPU, Varya can create a 5-second 720p video in just 45 seconds, a stark contrast to the 1,230 seconds needed by Wan 2.2. This efficiency is paired with an impressive pricing strategy, as Avataar plans to charge ₹0.48 ($0.005) per second of video, dramatically lower than competitors like Veo and Runway, which often charge $0.10 or more per second.
Rajan Anandan, Managing Director at Peak XV, emphasized that India is a video-centric market where video content is preferred over text. He noted that for AI video models to effectively reach a wide audience--including students, teachers, small businesses, and public services--cost reductions are essential for widespread adoption.
Furthermore, many existing video generation models often overlook cultural subtleties, producing generic outputs. Avataar AI has addressed this by training Varya with curated data that recognizes cultural elements such as cuisine, clothing, architecture, and festivals.
Varya will be available as an open-weight model on India's AI Kosh portal, the government's central repository for publicly accessible AI models and datasets. This allows developers to self-host or adapt the model according to their needs. Avataar also plans to collaborate with enterprise clients and video tool providers like Higgsfield and Adobe Firefly, enabling users to experiment with Varya on their website using text prompts or reference images.
Ultimately, Varya's launch signifies a pivotal moment in India's AI journey. Industry experts believe India has the potential to excel in AI by focusing on practical applications and fostering a robust developer ecosystem rather than competing solely on foundational models. The India AI Mission is part of a broader governmental strategy aimed at enhancing AI infrastructure, with aspirations to attract $200 billion in AI investments by 2028 and significantly increase GPU capacity in the near future.