How Avataar.ai Built an AI Video Model That Costs Just 48 Paise Per Second

The race to build frontier video-generation models has largely been dominated by companies in the United States and China. Now, Bengaluru-based Avataar.ai believes India can carve out its own place in the global market by making video AI dramatically cheaper, faster and more accessible than existing alternatives.

The company recently unveiled Varya, a video-generation model developed under the IndiaAI Mission. The model can generate video at approximately ₹ 0.48 per second, making it up to 10 times cheaper than several leading global video-generation models.

The company also plans to make Varya available through AIKosh, the IndiaAI Mission’s repository of AI models and datasets. Enterprises will be able to fine-tune the model using their own proprietary data.

.@Avataar launched #Varya today. Lowest cost video AI model in the world! This is a sneak peak of what Indian AI founders will do in every category of AI! @AluruSravanth @brettsingh @peakxvpartners pic.twitter.com/a7FwMnZxQk

— Rajan Anandan (@RajanAnandan) June 12, 2026

According to Sravanth Aluru, Co-Founder and CEO of Avataar.ai, the project stemmed from a simple observation that video is becoming the dominant medium of communication, yet the tools required to create it remain prohibitively expensive for a vast segment of Indian users.

Avataar.ai was co-founded by Aluru and Gaurav Baid in 2014. The company has raised more than $55 million from investors, including Tiger Global and Peak XV Partners, formerly Sequoia Capital India. It is also among the 12 startups selected under the IndiaAI Mission.

The company’s larger ambition is to bring more creators into the AI economy by making video-generation tools accessible to users who have traditionally been priced out of them.

“We really want to grow the pie,” Aluru says in an exclusive interview with AIM. “There’s an unserved, untapped market here because video is unaffordable for that segment.”

From Alibaba’s Wan to Varya

Rather than building a model entirely from scratch, Avataar.ai began with Wan 2.2, the open-source video model developed by Alibaba.

Aluru says the company’s goal was not merely to reduce costs but to preserve quality while improving efficiency.

The result, according to him, is a model that can generate more than 200 seconds of video for every ₹100 spent, making it substantially cheaper than many existing alternatives.

Central to Varya is a distillation technique that Aluru believes is often misunderstood.

Traditional distillation typically involves compressing a large model into a smaller one, sacrificing some capability in exchange for efficiency. Avataar.ai instead chose to retain the same 14-billion-parameter scale as the original model while redesigning the underlying diffusion process.

“We didn’t want to cheat-code it by just making the parameters low because quality would suffer,” Aluru explains.

The company sought to reduce the number of denoising steps required during video generation. While Wan 2.2 operates with 50 diffusion steps, Varya requires only four.

According to Aluru, three algorithmic innovations made that possible, including role-aware supervision, classifier-free guidance (CFG) augmentation, and a distillation framework inspired by MIT’s Distribution Matching Distillation (DMD) research.

“We actually started with the target of being three times cheaper,” he says. “We’ve ended up being 27 times cheaper.”

Aluru notes that generating a 720p video using the original model could take roughly 1,230 seconds. Varya can produce a similar output in approximately 45 seconds.

The company believes this near-real-time iteration will allow creators to experiment more freely, improving both productivity and creative outcomes.

Why India Needed Its Own Video Model

One of Varya’s most significant differentiators is its attempt to better understand Indian cultural nuances.

According to Aluru, global video models often generalise Indian concepts and fail to capture local context accurately.

Festivals, food, local environments and regional visual cues are often inaccurately represented by global models. To address this gap, Avataar.ai built more than 40,000 datasets designed to capture Indian cultural and regional nuances.

Aluru argues that video is not just another content format but a critical step toward building world models. He points to the work of Fei-Fei Li, whose startup World Labs is working on spatial intelligence and world models. The company is developing frontier AI systems that can perceive, generate and interact with the 3D world.

He notes that major AI breakthroughs over the past few years have centred on text and language models. Video, however, requires a different technological foundation built around diffusion models rather than the attention-based architectures that power LLMs.

“If India ever wants to get to AGI or world-model-level systems, you need to start understanding the world around you in much deeper granularity than just text,” Aluru says. 

From a market perspective, he believes the opportunity is equally compelling. Aluru points to users ranging from rural entrepreneurs and small businesses to teachers in government schools who could benefit from creating educational content but cannot afford existing video-generation tools.

He believes affordable video generation could help smaller creators compete with larger studios by compensating for limitations in equipment, lighting and production infrastructure.

At the same time, Aluru believes premium creators may also adopt the technology because of its speed.

“The creators who are creating premium content might actually say this is better because now I can iterate 20 times more,” he says.

Varya enters an increasingly competitive video AI market. China’s AI ecosystem is home to several prominent video generation models, including Alibaba’s Wan, Kuaishou’s Kling, Tencent’s HunyuanVideo, ByteDance’s Seedance, MiniMax’s Hailuo, Zhipu AI’s CogVideo, and ShengShu Technology’s Vidu.

Many of these models compete directly with leading US systems, including Google Veo and Runway. Notably, OpenAI shut down its video model Sora due to compute constraints.

Open Weights and the Sovereignty Debate

The decision to build on a Chinese open-source model could raise questions about technological sovereignty, particularly given India’s growing emphasis on indigenous AI capabilities.

Aluru does not see a contradiction. He notes that Wan 2.2 is distributed under an Apache licence and that Avataar.ai has made substantial foundational modifications to create Varya.

For Aluru, sovereignty is less about where the starting point came from and more about building capabilities and ownership within India. “I don’t think sovereignty really has a conflict with our ability to start doing smart innovation and frugal innovation,” he says.

The funding gap remains one of the biggest challenges facing Indian AI startups. “The dry powder of venture capital in India today is a minuscule portion of what’s happening in the US,” Aluru says. “Our entire dry powder might be less than the last round some of the names are raising in the US.”

That disparity, he argues, makes frugal innovation a strategic imperative for Indian AI companies.