Indian AI startup Avataar AI has launched its video generation model, Varya, targeting e-commerce and enterprise video production scenarios. The company states that the new model focuses on addressing two key challenges: generation speed and cost, and the ability to recognize India-specific cultural content.
Costs and speeds are reduced
Varya was not trained from scratch. Avataar AI is built on Alibaba's open-source video model Wan 2.2 and further compressed using distillation to create a lighter version tailored for its own business scenarios.
The company states that Varya’s generation process requires only 4 steps, while Wan 2.2 requires 50. On an NVIDIA H200 GPU, Varya generates a 5-second, 720p video in approximately 45 seconds, whereas Wan 2.2 takes 1,230 seconds.
Priced on a托管 service basis, Avataar AI plans to set Varya’s price at 0.48 Indian rupees per second of video, approximately $0.005. According to TechCrunch, this price is significantly lower than the typical rate of over $0.10 per second for similar products such as Veo, Kling, Luma, and Runway.
Mainly recognize India-local content
Image and video generation models often produce stereotypical or generalized content due to insufficient cultural context. Avataar AI states that Varya was trained on curated data to enhance its ability to recognize elements of Indian cuisine, attire, architecture, and festivals.
The company says users can currently try the model via the official website using text prompts or reference images. In addition to its own services, Avataar AI plans to open access to enterprise customers and is seeking partnerships with video tools such as Higgsfield and Adobe Firefly.
Will be released in an open-weight format
Varya will also be launched on the Indian government's AI Kosh platform in the form of an open-weight model, along with the accompanying training data. Developers can deploy the model themselves or modify it according to their needs.
This release also reflects the practical path of India’s AI industry. Rather than directly competing with the U.S., China, and other markets in developing foundational models, Indian industry players place greater emphasis on application deployment and developer ecosystem building. One reason is that local model development has long been constrained by insufficient computing power and limited access to high-quality data.
The Indian government previously launched the India AI Mission, with a budget of approximately $1.2 billion, providing subsidized computing power to selected startups and requiring them to publicly release their models. Avataar AI is one of the 12 companies selected. Earlier this year, the Indian government also stated its goal of attracting $200 billion in AI investment by 2028 and more than doubling GPU capacity within six months.