Model Distillation for Efficiency

Avataar AI’s new video model, Varya, demonstrates the effectiveness of model distillation in resource-constrained environments. Rather than training from scratch, the company compressed Alibaba’s Wan 2.2 model. By reducing the inference process from 50 steps to just 4, the team achieved a 10x increase in generation speed. On an NVIDIA H200 GPU, Varya generates a 5-second 720p clip in 45 seconds, compared to 1,230 seconds for the base model.

Cultural Context and Economic Accessibility

A primary barrier to AI adoption in India is the lack of cultural nuance in Western-trained models, which often produce stereotypical outputs. Avataar addressed this by training Varya on a curated dataset specifically designed to recognize Indian architecture, clothing, food, and festivals. Furthermore, the company is targeting mass-market adoption by pricing its service at ₹0.48 ($0.005) per second—roughly 20 times cheaper than current global leaders like Luma, Runway, and Kling. This pricing strategy is intended to make video AI viable for students, MSMEs, and public services in a video-first market.

Strategic Alignment with India’s AI Mission

Varya is a product of the India AI Mission, a $1.2 billion government initiative that provides startups with subsidized compute in exchange for releasing models publicly. Avataar is releasing Varya as an open-weight model on the India AI Kosh portal, allowing developers to self-host or fine-tune the model. This launch highlights a broader shift in India’s AI strategy: prioritizing the development of localized applications and a robust developer ecosystem over attempting to compete directly with global foundation model labs.