India’s production of AI models is lagging compared to the US, Europe, and China. Only a few startups have released models, most of which are large-scale language or speech models. To spur further development, the government has launched the India AI Mission, an approximately $1.2 billion initiative. This, among other things, gives selected startups access to subsidized GPU computing in exchange for releasing their models to the public. Avataar AI, one of the 12 startups selected for the program, announced a new video model called Varya built to understand local context, including identifying different festivals, foods, and clothing.
The Peak XV-backed startup focuses on creating video tools for e-commerce and did not build Varya from scratch. It started with Wan 2.2, a publicly available video generation model released by Alibaba that used a technique called distillation. This essentially compressed the functionality of the model into a leaner, faster version that was optimized for Avataar’s specific use case. The result is a model that runs in 4 steps instead of 50 in Wan 2.2, producing videos 10 times faster and at a fraction of the cost.
Specifically, with the NVIDIA H200 GPU, Varya can generate a 5-second 720p clip in 45 seconds, compared to 1,230 seconds with Wan 2.2.
The most impressive aspect of Varya may be its price. The company plans to charge ₹0.48 ($0.005) per second for videos on its hosted services. This is much cheaper than models like Veo, Kling, Luma, and Runway, which typically cost $0.10 per second or more. The price difference is approximately 20 times.
“India is a video-first market. This can be seen in every mass consumer internet product in India. Video trumps text. Current AI video models are too expensive for use at population scale in India. If video AI is to reach students, teachers, MSMEs, creators, enterprises, and public services, costs will have to come down dramatically. Cost is the biggest key to AI adoption in India.” Peak Rajan Anandan, managing director of XV, told TechCrunch.
Image and video generation models often miss cultural nuances and produce stereotypical or generic output, an issue TechCrunch previously reported on. Avatar AI says it has trained Varya to recognize cultural nuances such as food, clothing, architecture, and festivals using carefully selected data.
Varya will be released as an openweight model on India’s AI Kosh portal, the Indian government’s central repository of publicly available AI models and datasets, along with training data. This means developers can self-host it or modify it to suit their needs. Avatar also plans to offer this model to enterprise customers and is open to partnering with video tools such as Higgsfield and Adobe Firefly. Anyone can try it out now on our website using text prompts and reference images.
Varya’s launch reflects a fundamental trade-off in India’s AI ambitions. Industry veterans point out that India can make its mark in the AI space by creating applications and a robust developer ecosystem rather than competing on foundational models. And there is a reason for that realism. India’s model development lags behind its global rivals due to a lack of computing and limited quality data available.
The India AI Mission is also part of a broader government effort to bridge that gap. Last year, the company selected 12 startups to develop AI models, including Avatar AI, to provide cost-effective computing. Earlier this year, IT Minister Ashwini Vaishnaw said India aims to attract $200 billion in AI investment by 2028 and more than double its GPU capacity within six months.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
