India has lagged global rivals in building cutting-edge AI models, constrained by scarce compute and fragmented data. While the U.S., Europe, and China race ahead with large language and multimodal systems, only a handful of Indian startups have shipped serious models. The government’s India AI Mission, a roughly $1.2 billion program that offers subsidized GPU access in exchange for open releases, is designed to change that trajectory.

Among the 12 startups selected for this mission, Bengaluru-based Avataar AI is emerging as a test case for whether India can build AI that is both globally competitive and locally grounded. Known for its video tools for e-commerce, the company has unveiled Varya, a video generation model tuned to India’s cultural and economic realities.

Instead of training from scratch, Avataar began with Wan 2.2, a publicly released video model from Alibaba. Using distillation, the team compressed Wan’s capabilities into a smaller, faster system optimized for their workloads. Where Wan 2.2 needs 50 inference steps, Varya runs in four, cutting both latency and cost.

On an NVIDIA H200 GPU, Varya can generate a five-second 720p clip in about 45 seconds, compared with more than 20 minutes for the original Wan 2.2. That performance shift is central to Avataar’s pitch: video AI that can operate at India’s population scale rather than as a luxury tool for a few well-funded creators.

Pricing underscores that ambition. Avataar plans to charge ₹0.48 per second of video on its hosted service, under one U.S. cent. Competing systems such as Veo, Kling, Luma, and Runway typically cost at least 20 times more, putting them out of reach for many Indian students, small businesses, and regional creators.

Equally important is cultural fluency. Global image and video models often misinterpret Indian festivals, clothing, food, and architecture, defaulting to generic or Westernized imagery. Avataar says it has trained Varya on curated datasets to better recognize local contexts, from Diwali and Onam to street food, saris, and vernacular design.

Varya will be released as an open-weight model on the government’s AI Kosh portal, along with its training data, allowing developers to self-host, fine-tune, or embed it into their own products. Avataar will also integrate Varya into its enterprise offerings and is exploring partnerships with established creative tools.

For policymakers, Varya is more than a single product launch. It is an early signal that India’s AI strategy may hinge less on building the biggest foundation models and more on delivering affordable, culturally aware systems that millions can actually use.