原文 · 未翻译
India’s AI model output has been slow compared to the U.S., Europe, and China. Only a few startups are releasing models, and most of them are large language models or voice models. To encourage more development, the government launched the India AI Mission, a roughly $1.2 billion initiative that — among other things — gives selected startups access to subsidized GPU compute in exchange for releasing their models publicly. One of the 12 startups selected for the program, Avataar AI, has launched a new video model called Varya that is built to understand local context — such as identifying different festivals, food, and clothing.
The Peak XV-backed startup, which focuses on creating video tools for e-commerce, didn’t build Varya from scratch. It started with Wan 2.2, a publicly available video generation model released by Alibaba, and used a technique called distillation — essentially compressing the model’s capabilities into a leaner, faster version optimized for Avataar’s specific use cases. The result is a model that runs in four steps rather than Wan 2.2’s 50, producing video 10 times faster and at a fraction of the cost.
To put that in concrete terms: Using an Nvidia H200 GPU, Varya can generate a five-second 720p clip in 45 seconds, compared to 1,230 seconds for Wan 2.2.
The most striking aspect of Varya may be its price. The company plans to charge ₹0.48 ($0.005) per second of video on its hosted service — far cheaper than models like Veo, Kling, Luma, and Runway, which typically charge $0.10 or more per second. That’s a roughly 20x price difference.
“India is a video-first market. We see this across every large consumer internet product in India: video wins over text. 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,” Peak XV’s managing director Rajan Anandan told TechCrunch.
Image and video generation models often miss cultural nuances and produce stereotyped or generic outputs — a problem TechCrunch has reported on before. Avataar AI says it has used curated data to train Varya to recognize cultural nuances including food, clothing, architecture, and festivals.
Varya will be released as an open-weight model on India’s AIKosh portal — the Indian government’s centralized repository for publicly available AI models and datasets — along with its training data, meaning developers can self-host or modify it for their own needs. Avataar also plans to make the model available to its enterprise customers and says it is open to partnerships with video tools, including Higgsfield and Adobe Firefly. Anyone can try it now on its website using text prompts or reference images.
Varya’s launch reflects a fundamental tradeoff in India’s AI ambitions. Industry veterans have noted that India can make its mark in AI by creating applications and a robust developer ecosystem rather than competing on foundation models. And there’s a reason for that pragmatism: Model development has been slower in India than in global rivals due to a lack of compute and limited quality data availability.