Avataar AI Launches Varya: A Cheap, Fast, and Culturally Aware Video Model for India
Avataar AI, a startup selected for India's AI Mission, launched Varya, a video generation model that is 10x faster and 20x cheaper than rivals, with cultural understanding.
Varya: Built for India's Scale
Avataar AI, one of 12 startups chosen for India's $1.2 billion AI Mission, has unveiled a new video generation model called Varya. The model is designed to address India's unique needs: it is much cheaper and faster than existing alternatives, and it understands local cultural contexts such as festivals, food, and clothing.
Varya is not built from scratch. It is based on Alibaba's open-source Wan 2.2 model, but Avataar applied a technique called distillation to compress the model, reducing the number of steps from 50 to 4. This results in video generation that is ten times faster and significantly cheaper. On an Nvidia H200 GPU, Varya can produce a five-second 720p clip in 45 seconds, compared to 1,230 seconds for the original model. Pricing is set at ₹0.48 ($0.005) per second of video, roughly 20 times lower than competitors like Veo, Kling, Luma, and Runway, which charge $0.10 or more per second.
Rajan Anandan, managing director at Peak XV (which backs Avataar), noted that cost is the biggest barrier to AI adoption in India, especially for a video-first market. To overcome cultural biases common in AI models, Avataar trained Varya with curated data that recognizes Indian nuances.
The model will be released as open-weight on India's AIKosh portal, along with its training data, allowing developers to self-host or modify it. Avataar also plans to offer Varya to enterprise customers and is open to partnerships with tools like Higgsfield and Adobe Firefly. Users can try it now via text prompts or reference images on the company's website.
Varya's launch reflects a pragmatic approach: India may not lead in foundation models but can excel in applications and ecosystems, backed by government initiatives to close the compute gap.
