Avataar's open-weight Varya video model launches for India at 20x lower cost

Avataar AI, one of 12 startups selected for India's ~$1.2 billion India AI Mission, launched Varya, a video-generation model built specifically for India's scale and cultural context — able to identify local festivals, food, and clothing. Rather than building from scratch, Avataar distilled Alibaba's publicly available Wan 2.2 model into a leaner version that runs in just 4 steps instead of 50, producing video 10x faster.

The performance and price numbers are striking: on an NVIDIA H200, Varya generates a 5-second 720p clip in 45 seconds versus 1,230 seconds for Wan 2.2. The company plans to charge ₹0.48 ($0.005) per second on its hosted service — roughly a 20x price reduction versus Veo, Kling, Luma, and Runway, which typically charge $0.10 or more per second. Peak XV's Rajan Anandan called cost 'the biggest unlock for AI adoption in India.'

Varya will be released as an open-weight model on India's AI Kosh portal along with its training data, letting developers self-host or modify it, and will also reach Avataar's enterprise e-commerce customers. The launch is a notable data point in the global push for cheaper, regionally-aware, open models — echoing the week's open-source-vs-closed debate and the cost pressures squeezing frontier providers. The open question is quality: a 4-step distilled model trades some fidelity for speed, and whether it satisfies population-scale Indian use cases remains to be proven in the wild.