Avataar AI, an Indian startup backed by Peak XV, has unveiled a new video-generation model called Varya, designed to understand Indian cultural contexts while offering lower costs and faster processing.
The company, which develops video tools for e-commerce, is one of the 12 startups selected under the Indian government’s India AI Mission, a roughly $1.2 billion initiative aimed at promoting domestic AI development.
Built on Alibaba’s Wan 2.2 model
Avataar AI did not build Varya from scratch.
The startup used Wan 2.2, an open video-generation model released by Alibaba, and applied a distillation technique to create a lighter and faster version optimized for its use cases.
According to the company, Varya completes video generation in four steps, compared with 50 steps for Wan 2.2.
Faster video generation
Using an Nvidia H200 GPU, the company said Varya can generate:
- A five-second 720p video in 45 seconds.
By comparison, Wan 2.2 requires:
- 1,230 seconds to produce a similar clip.
Avataar said the model delivers videos about 10 times faster while requiring significantly lower computing resources.
Lower pricing
The company plans to charge â‚ą0.48 ($0.005) per second of generated video through its hosted service.
According to the report, competing platforms including:
- Veo
- Kling
- Luma
- Runway
typically charge $0.10 or more per second.
That represents a cost difference of roughly 20 times.
Designed for Indian cultural contexts
Avataar AI said Varya has been trained with curated datasets to recognize local nuances, including:
- Food.
- Clothing.
- Architecture.
- Festivals.
The company said existing image and video models often struggle with cultural representation and may generate stereotypical outputs.
Open-weight release planned
Varya will be made available as an open-weight model through AIKosh, the Indian government’s repository for AI models and datasets.
Avataar also plans to release the training data.
Developers will be able to:
- Self-host the model.
- Modify it for specific applications.
- Integrate it into their own tools.
The startup said it is open to partnerships with companies including:
- Higgsfield
- Adobe Firefly
The model is already accessible through Avataar’s website using text prompts or reference images.
Backed by India’s AI push
The launch comes amid efforts by the Indian government to strengthen the country’s AI ecosystem.
Through the India AI Mission, selected startups receive subsidized access to GPU computing resources in exchange for making their models publicly available.
Earlier this year, Information Technology Minister Ashwini Vaishnaw said India aims to attract $200 billion in AI investment by 2028 and plans to more than double the country’s GPU capacity within six months.
Focus on applications
Industry executives have argued that India can play a larger role in AI by developing applications and supporting a strong developer ecosystem instead of competing directly in foundation models.
According to the report, limited access to computing infrastructure and high-quality datasets has slowed model development in India compared with the United States, Europe and China.
Investor view
“India is a video-first market. We see this across every large consumer internet product in India: video wins over text,” Rajan Anandan, managing director at Peak XV,
“Current AI video models are too expensive for population-scale use in India. If video AI is going to reach students, teachers, MSMEs, creators, enterprises, and public services, costs have to come down dramatically. Cost is the biggest unlock for AI adoption in India,” he said.
TL;DR
Avataar AI has launched Varya, a low-cost video-generation model optimized for Indian users. Built on Alibaba’s Wan 2.2 model, it offers faster generation speeds, lower costs and support for Indian cultural contexts.
AI summary
- Avataar AI launched the video model Varya.
- The model is based on Alibaba’s Wan 2.2.
- Varya generates videos faster and at lower cost.
- It has been trained to recognize Indian cultural nuances.
- The model will be released on the AIKosh platform.
