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

Overview

Indian startup Avataar has launched Varya, a text-to-video AI model that generates clips for about 0.48 rupees, or roughly $0.005, per second. The company says this is around 20 times cheaper than global tools such as Veo, Kling, Luma, and Runway. Varya was built by distilling Alibaba's open Wan 2.2 model down to 4 generation steps from 50, letting it run about 10 times faster. It is one of 12 startups selected for India's roughly $1.2 billion AI Mission.

Key Takeaways

Avataar's Varya generates AI video for about 0.48 rupees ($0.005) per second, which the company positions as up to 20 times cheaper than leading rivals.
The model was created by distilling Alibaba's open Wan 2.2 video model, cutting generation to 4 steps from 50 and running about 10 times faster.
A 5-second 720p clip takes about 45 seconds to produce on an NVIDIA H200 GPU, compared with roughly 1,230 seconds for the original Wan 2.2 model.
Varya is trained to recognize Indian cultural details such as festivals, food, clothing, facial features, streetscapes, and architecture that global models often miss.
Avataar will release Varya as an open-weight model with its training data on India's AIKosh portal, the government's public repository for AI models and datasets.
The company is one of 12 startups chosen for the IndiaAI Mission, a government program worth around $1.2 billion, and built Varya using subsidized public compute.

Stats & Key Facts

#Pricing of about 0.48 rupees, or $0.005, per second of generated video.
#Roughly 20 times cheaper than rivals that charge upward of $0.10 per second.
#Generation cut to 4 steps from 50, producing about 10 times faster output.
#A 5-second 720p clip renders in about 45 seconds versus roughly 1,230 seconds for Wan 2.2.
#One of 12 startups selected for the roughly $1.2 billion IndiaAI Mission.
#Avataar raised $45 million in a 2022 Series B round from investors including Tiger Global and Peak XV.

Varya Prices AI Video at About $0.005 Per Second

Cost is the central bet behind the launch.

Avataar built Varya to generate video for about 0.48 rupees, or roughly half a US cent, per second. The company says that figure is around 20 times lower than global tools such as Veo, Kling, Luma, and Runway, which it pegs at upward of $0.10 per second.

The reasoning is simple. To reach a population the size of India's, the per-clip price has to fall to a level businesses and everyday creators afford at scale. Avataar treats low cost not as a discount feature but as the condition for mass adoption in markets where premium tools stay out of reach.

How Distilling Wan 2.2 Made Varya 10 Times Faster

The speed gain comes from a technical shortcut, not a brand-new model.

›Varya is built on Alibaba's publicly available Wan 2.2 video model using a distillation technique.
›Distillation cut the generation process to 4 steps from the original 50 steps.
›The shorter path lets Varya run about 10 times faster than Wan 2.2 while keeping comparable output quality.
›Avataar describes Varya as India's first distilled video AI model.

5-Second 720p Clip in 45 Seconds on an H200 GPU

The speed claim has a concrete benchmark behind it.

Avataar reports that a 5-second 720p clip takes about 45 seconds to generate on an NVIDIA H200 GPU. The same clip on the original Wan 2.2 model takes roughly 1,230 seconds, the company says.

That gap is the practical payoff of the distillation work. Faster rendering on the same hardware lowers compute cost per clip, which feeds directly into the low per-second price the company is promoting.

Trained for Indian Festivals, Food, Clothing, and Streetscapes

Cultural accuracy is the second pillar of the pitch.

Varya is trained to recognize Indian context that mainstream models often get wrong, including festivals, food, clothing, facial features, streetscapes, and architecture. Avataar frames this as a fix for the stereotyping and accuracy problems common when global tools generate images and video for non-Western settings.

Co-founder and chief executive Sravanth Aluru said leading video models are expensive to run and tend to under-represent India's cultural, linguistic, and visual diversity. Building local context into the model is meant to make output usable for Indian businesses without heavy correction.

Backed by the $1.2 Billion IndiaAI Mission

Public infrastructure underpins the project.

›Avataar is one of 12 startups selected for the IndiaAI Mission, a government program worth around $1.2 billion.
›The company built Varya using subsidized compute provided through the program.
›In exchange for that support, Avataar plans to release Varya publicly as an open-weight model.
›The launch was attended by S Krishnan, secretary at the Ministry of Electronics and Information Technology, who called it a milestone for indigenous foundational AI models.

Open-Weight Release on AIKosh With Training Data

Avataar plans to share the model rather than keep it closed.

Avataar will publish Varya as an open-weight model on AIKosh, India's centralized government portal for publicly available AI models and datasets. The release is set to include the training data, so other developers build on the model.

Alongside the open release, the company is opening access through a public beta where users generate clips from text prompts or reference images, and it points to potential partnerships with tools such as Adobe Firefly and Higgsfield. The approach fits India's broader strategy of competing on applications and developer ecosystems rather than racing to build the largest foundation models.

Avataar's Background and Investors

The startup is not new to AI media tools.

›Avataar was founded in 2014 and is led by chief executive and co-founder Sravanth Aluru.
›The company raised $45 million in a Series B round in 2022 from investors including Tiger Global and Sequoia Capital India, now known as Peak XV.
›Peak XV managing director Rajan Anandan said cost is the biggest barrier to AI adoption in India and current models are too expensive for population-scale use.

Frequently Asked Questions

How much does Avataar's Varya cost to generate video?

Varya generates video for about 0.48 rupees, or roughly $0.005, per second. Avataar says that is around 20 times cheaper than rivals like Veo, Kling, Luma, and Runway, which charge upward of $0.10 per second.

What model is Varya based on?

Varya is built by distilling Alibaba's publicly available Wan 2.2 video model. The distillation cut the generation process to 4 steps from 50, letting Varya run about 10 times faster while keeping comparable quality.

Why is Varya described as culturally aware?

It is trained to recognize Indian context such as festivals, food, clothing, facial features, streetscapes, and architecture. Avataar built this in to address the stereotyping and accuracy issues common when global models generate content for non-Western settings.

What is the IndiaAI Mission and how is Avataar involved?

The IndiaAI Mission is a government program worth around $1.2 billion to support local AI development. Avataar is one of 12 startups selected, and it built Varya using subsidized public compute provided through the program.

Will Varya be open to other developers?

Yes. Avataar plans to release Varya as an open-weight model on India's AIKosh portal, including the training data, so other developers build on it.

Varya reflects India's strategy of using public infrastructure to help local startups build affordable, culturally accurate AI. By cutting per-second cost and rendering time sharply, Avataar is betting that price, not model size, is what opens AI video to population-scale use.