India’s Avataar AI unveils video model at $0.005/sec, 27x cheaper than rivals
▼ Summary
– Avataar AI launched Varya, an open-weight video model costing $0.005 per second, which is 27 times cheaper than comparable models.
– Varya achieves cost savings through distillation, compressing Alibaba’s Wan 2.2 model to run in four steps instead of 50 for faster generation.
– The model prioritizes affordability and accessibility over quality, targeting India’s price-sensitive market rather than competing with frontier models.
– Varya is trained on curated data to accurately render Indian culture, including clothing, food, architecture, and festivals, unlike Western-trained models.
– Avataar is part of India’s $1.2 billion AI Mission, and Varya will be released on the government’s AIKosh portal as an open-weight model.
Bangalore-based Avataar AI has introduced Varya, one of India’s first indigenous video generation models, priced at roughly $0.005 per second , or about 0.48 rupees. Founder Sravanth Aluru, a former Deutsche Bank investment banker and an IIT Mumbai and Microsoft alum, claims this makes Varya 27 times cheaper than comparable open-source video models currently available.
That dramatic cost reduction is achieved through distillation. Avataar built on Alibaba’s Wan 2.2, a publicly accessible video model, and compressed its core capabilities into a streamlined version that operates in just four steps instead of the usual 50. The result is generation speeds ten times faster at a fraction of the cost. For context, competitors like Veo, Kling, Luma, and Runway typically charge $0.10 or more per second.
Varya does not aim to match the raw quality of frontier models from the US or China. ByteDance’s Seedance, Kuaishou’s Kling, and Alibaba’s Wan have pushed motion realism and audio generation well beyond what Varya currently offers. Instead, the pitch is about scale and accessibility in a market of 1.4 billion people, where cost competitiveness often outweighs peak performance.
What truly sets Varya apart is its cultural specificity. Rather than adapting a Western-trained model, Avataar used curated datasets to train Varya to accurately render Indian clothing, food, architecture, festivals, and everyday settings. Global models, primarily trained on Western data, consistently fail at this, producing culturally inaccurate outputs that limit their usefulness for Indian businesses, education, and public services.
The model is open-weight and will be hosted on India’s AIKosh portal, the government’s centralized repository for AI models and datasets. Avataar is one of 12 startups selected under the IndiaAI Mission, a roughly $1.2 billion initiative that provides subsidized GPU compute in exchange for public release of models.
Avataar has raised $55 million from Peak XV Partners and Tiger Global. Originally focused on video tools for e-commerce, Varya is the company’s first foundation model. It reflects a broader shift among Indian startups toward building sovereign AI rather than simply renting Western infrastructure. Earlier this year, Sarvam and BharatGen also launched their own foundational models under the same program.
India’s AI strategy differs from Europe’s or China’s. The goal is not to build the biggest model, but to develop models that function for its population at a price its market can absorb. At $0.005 per second, Varya is testing whether an affordable, culturally relevant video model can gain traction faster than a technically superior but expensive Western alternative. In a country where AI startups are already building for local needs at scale, the answer may well be yes.
(Source: The Next Web)