India’s artificial intelligence (AI) landscape has been evolving, albeit at a slower pace compared to regions like the U.S., Europe, and China. To accelerate this growth, the Indian government initiated the India AI Mission, a substantial $1.2 billion program aimed at fostering AI development. This initiative provides selected startups with subsidized GPU compute resources, encouraging them to release their models publicly.
Among the 12 startups chosen for this program is Avataar AI, a company specializing in video tools for e-commerce. Avataar has recently unveiled Varya, a new video generation model designed to understand and incorporate local cultural contexts, such as recognizing various festivals, cuisines, and traditional attire.
Rather than building Varya from the ground up, Avataar AI utilized Wan 2.2, a publicly available video generation model developed by Alibaba. Through a process known as distillation, they compressed the model’s capabilities into a more efficient version tailored to their specific applications. This optimization allows Varya to operate in four steps, a significant reduction from Wan 2.2’s 50 steps, resulting in video production that is ten times faster and considerably more cost-effective.
To illustrate this efficiency, using an NVIDIA H200 GPU, Varya can generate a 5-second 720p video clip in just 45 seconds, compared to the 1,230 seconds required by Wan 2.2.
One of Varya’s most compelling features is its affordability. Avataar plans to charge ₹0.48 ($0.005) per second of video on its hosted service. This pricing is substantially lower than that of other models like Veo, Kling, Luma, and Runway, which typically charge $0.10 or more per second—a difference of approximately 20 times.
Rajan Anandan, Managing Director at Peak XV, emphasized the significance of this development, stating that India is a video-centric market where video content consistently outperforms text across major consumer internet platforms. He noted that existing AI video models are prohibitively expensive for widespread use in India. To make video AI accessible to students, educators, small and medium-sized enterprises (SMEs), content creators, corporations, and public services, costs must be significantly reduced. Anandan highlighted that cost reduction is the primary catalyst for AI adoption in India.
AI models for image and video generation often struggle with cultural nuances, leading to outputs that are stereotypical or generic. Avataar AI addressed this issue by training Varya with curated data to accurately recognize and represent cultural elements such as food, clothing, architecture, and festivals.
Varya will be released as an open-weight model on India’s AI Kosh portal, the government’s centralized repository for publicly available AI models and datasets. This means developers can self-host or modify the model to suit their specific needs. Additionally, Avataar plans to offer Varya to its enterprise clients and is open to collaborations with video tools like Higgsfield and Adobe Firefly. The model is currently accessible on Avataar’s website, where users can generate videos using text prompts or reference images.
The launch of Varya underscores a strategic shift in India’s AI ambitions. Industry experts have suggested that India’s strength lies in developing AI applications and fostering a robust developer ecosystem, rather than competing directly in the creation of foundational models. This pragmatic approach acknowledges the slower pace of model development in India compared to global counterparts.
By introducing Varya, Avataar AI not only showcases India’s potential in AI innovation but also addresses the critical need for cost-effective, culturally aware AI solutions tailored to the Indian market. This development could serve as a catalyst for broader AI adoption across various sectors in India, paving the way for more inclusive and accessible technological advancements.
