Cheaper, faster, and culturally aware, Avataar’s video AI is built for India’s scale

AI model production in India is slow compared to the US, Europe and China. Only a few startups are releasing models, and most of them are large language or voice models. To encourage more development, the government launched the India AI Missiona roughly $1.2 billion initiative that, among other things, gives select startups access to subsidized GPU computing in exchange for publicly releasing their models. One of the twelve startups selected for the program, Avatar AIhas launched a new video model called Varya, which is built to understand local context, such as identifying different festivals, food and clothing.

The Peak XV-backed startup, which focuses on creating video tools for e-commerce, didn’t build Varya from the ground up. It started with Wan 2.2, a publicly available video generation model released by Alibaba, and used a technique called distillation, which essentially compressed the model’s capabilities into a leaner, faster version optimized for Avatar’s specific use cases. The result is a model that works in four steps instead of Wan 2.2’s 50, producing video 10 times faster and at a fraction of the cost.

To make that concrete, using an NVIDIA H200 GPU, Varya can generate a 5-second 720p clip in 45 seconds, compared to 1,230 seconds for Wan 2.2.

Perhaps the most notable aspect of Varya is its price. The company plans to charge ₹0.48 ($0.005) per second of video on its hosted service – much cheaper than models like Veo, Kling, Luma and Runway, which typically charge $0.10 or more per second. That’s a price difference of about 20x.

“India is a video-first market. We see this with every major consumer internet product in India: video is beating text. Current AI video models are too expensive for population-scale use in India. For video AI to reach students, educators, MSMEs, makers, enterprises and public services, costs need to drop dramatically. Cost is the biggest unlock for AI adoption in India,” Rajan Anandan, director of Peak XV, told TechCrunch.

Image and video generation models often miss cultural nuances and produce stereotypical or generic results – a problem TechCrunch has previously reported on. Avatar AI says it used curated data to train Varya to recognize cultural nuances, including food, clothing, architecture and festivals.

Varya will be released as an open-weight model India’s AI Kosh Portal – the Indian government’s centralized repository for publicly available AI models and datasets – along with the training data, meaning developers can self-host it or customize it to their own needs. Avatar also plans to make the model available to its enterprise customers and says it is open to partnerships with video tools like Higgsfield and Adobe Firefly. Anyone can try it now on the website using text prompts or reference images.

The launch of Varya reflects a fundamental trade-off in India’s AI ambitions. Industry veterans have noted that India can make its mark in AI by create applications and a robust one developer ecosystem instead of competing on foundation models. And there’s a reason for that pragmatism: model development has been slower in India than in global rivals, due to a lack of computing power and the limited availability of quality data.

The India AI mission is also part of a broader government approach to closing that gap. Last year it selected twelve startups, including Avatar AI, to develop AI models and provide them with cost-efficient computing power. Earlier this year, IT Minister Ashwini Vaishnaw said India aims to attract $200 billion in AI investments by 2028. double the GPU capacity within six months.