Every major AI company is racing to build the biggest model. India just decided to build the most useful one.
There is a teacher in a government school in Rajasthan who wants to explain photosynthesis to her students using a short video.
There is a saree seller in Surat who wants a thirty-second product video for his Instagram page but cannot afford a videographer. There is a municipal corporation in a small town that wants to communicate a public health message in the local language without hiring a production house.
None of them can afford the AI video tools that Silicon Valley has been celebrating for the past two years. That is the problem Varya is trying to solve.
What Is Varya and Who Built It
On June 12, 2026, the Ministry of Electronics and Information Technology launched Varya, India’s first distilled AI video model, at an event in New Delhi attended by MeitY Secretary S. Krishnan.
The model was built by Bengaluru-based startup Avataar AI, which was among twelve companies selected under the IndiaAI Mission, the Indian government’s roughly 1.2 billion dollar national initiative to build indigenous artificial intelligence capabilities.
Avataar AI is backed by Peak XV, formerly Sequoia India, and has spent the last several years building video tools specifically for e-commerce. They did not build Varya from scratch.
They started with Wan 2.2, a publicly available video generation model released by Alibaba, and used a technique called distillation to compress it into something leaner, faster, and significantly cheaper to run.
The result is a 14-billion-parameter model that can generate a five-second video clip in 45 seconds on an Nvidia H200 GPU. The same task takes Wan 2.2, the base model it was built from, over 1,200 seconds. That is not a small improvement. That is a fundamentally different experience.
How Much Varya Cost?
Varya costs Rs 0.48 per second of video generated. That is less than fifty paise. That is roughly half a rupee.
To understand why that number matters, you need to know what the alternatives cost. Google’s Veo 3.1, currently considered one of the best AI video models in the world, starts at $0.15 per second in fast mode.
That is approximately Rs 12.50 per second at current exchange rates. Runway, Kling, and Luma, three of the most widely used global AI video platforms, typically charge $0.10 or more per second, which works out to roughly Rs 8 to Rs 9 per second.
Varya charges Rs 0.48. TechCrunch described the price difference as roughly 20 times cheaper than models like Veo, Kling, Luma, and Runway.
Avataar’s own internal benchmarks describe it as up to 10 times more cost-efficient than several leading global models. The difference in those two numbers comes down to which competitors you compare against and how you measure efficiency. What is not in dispute is that Varya is dramatically cheaper than anything currently available at comparable quality.
One important thing to note: the efficiency and cost comparisons are based on Avataar’s internal benchmarks, not an independently verified third-party test. That does not make the numbers wrong, but it is worth keeping in mind as the model moves from launch event to real-world use.
How Varya Actually Works
The workflow is simple by design. You type an idea. Varya generates a video. You can extend that video into a longer multi-clip story.
The technical innovation behind this simplicity is the distillation technique. Most AI video models generate video through a process that involves roughly 50 computational steps.
Each step takes time and costs money. Varya reduces those 50 steps to 4, while maintaining output quality that Avataar says is comparable to leading global models including Wan 2.2 and Gemini Veo 3.
Think of it like this. If a traditional AI video model is a full commercial kitchen that needs twenty chefs, premium equipment, and three hours to produce a meal, Varya is a well-designed smaller kitchen that gets you the same meal in twenty minutes with a fraction of the resources.
The model is also being released as an open-weight model on India’s AI Kosh portal at aikosh.indiaai.gov.in. This means developers can download it, run it on their own infrastructure, modify it, and build products on top of it without paying Avataar anything. The Rs 0.48 per second pricing applies only to Avataar’s hosted service, which is the ready-to-use version for non-technical users.
The Part Nobody Is Talking About
Every launch announcement leads with the price and the speed. Those are important. But the detail that deserves more attention is the cultural training.
Global AI video models are built primarily on Western visual data. When you ask Sora or Runway to generate a video of a festival, you are likely to get something that looks generically festive. When you ask it to generate a video of a street market, you are likely to get something that looks vaguely European or American.
Varya has been specifically trained to understand and generate Indian visual contexts. Diwali looks like Diwali. A Tamil wedding looks like a Tamil wedding. A vegetable market in Lucknow looks like a vegetable market in Lucknow, not a farmers’ market in California.
This cultural specificity is not a cosmetic feature. It is the difference between a tool that technically works and a tool that is actually useful for the 1.4 billion people it is supposed to serve. An MSME owner in Nagpur making a product video does not want generic global aesthetics. They want something that looks like their world and speaks to their customer.
Sravanth Aluru, CEO and co-founder of Avataar, put it plainly at the launch: “For a country of 1.4 billion people, affordability is not a feature. It is a prerequisite.”
What This Fits Into
Varya did not arrive in isolation. It is the first publicly visible output of a much larger national strategy.
The IndiaAI Mission was set up specifically to prevent India from being permanently dependent on foreign AI infrastructure.
It provides selected startups with access to subsidised GPU compute, which is the expensive computing power required to train and run large AI models, in exchange for releasing their models publicly. Varya is the first video model to come out of this programme.
India has done this before with technology. UPI did not just build a payments system. It built a payments system designed for Indian conditions, Indian scale, and Indian economics, and it ended up processing more transactions than Visa in its peak months.
ISRO’s Mars mission cost less than the Hollywood film Gravity. The instinct to build world-class capability at a fraction of global cost is not new in India. Varya is that instinct applied to artificial intelligence.
What Comes Next
Varya is available to try right now at varya.avataar.ai. The developer API and the AI Kosh portal listing are still being rolled out. Avataar has said it is open to partnerships with video platforms including Adobe Firefly, and is targeting enterprise customers alongside individual creators and MSMEs.
The model will be tested hardest not at launch events but in the months that follow, when real users in real conditions put it to work and compare the output honestly against what the global alternatives produce.
The price is real. The cultural training is real. The speed improvement is real. Whether the output quality truly holds up against Veo and Runway at scale is the question that will determine whether Varya becomes a genuine tool for a billion people or a well-intentioned proof of concept that gets quietly forgotten.
India has built the model. The harder work starts now.
