📰 Key Highlights

Indian AI startup Avataar AI has launched a video generation model named Varya, designed for the Indian cultural context, capable of recognizing local festivals, clothing, food and architecture. Avataar AI is one of twelve selected startups in the Indian government’s “India AI Mission”, a program with approximately $1.2 billion in total scale, offering GPU compute subsidies in exchange for model release.

Varya wasn’t trained from scratch—it was built on Alibaba’s open-source Wan 2.2 video generation model, compressed through knowledge distillation, reducing steps from 50 to just 4. That’s a 10x speed boost and major cost drop. The numbers are pretty wild: generating a 5-second 720p video with an NVIDIA H200 GPU takes Varya just 45 seconds, while the original Wan 2.2 needs 1,230 seconds.

Pricing is where Varya really stands out. Avataar AI plans to offer it at ₹0.48 (about $0.005) per second of video on their managed service—compared to mainstream models like Veo, Kling, Luma, and Runway charging $0.10+ per second, that’s roughly a 20x price difference. Peak XV Managing Director Rajan Anandan pointed out that India is a video-first market, but existing AI video models are too expensive for widespread adoption—they need to drastically cut costs to reach students, teachers, SMBs, and public services.

Varya will be released as an open-weight model on the Indian government’s AI Kosh platform, complete with training data, so developers can deploy or modify it themselves. Right now, anyone can try it directly on the website with text or reference images.

💬 JudyAI Lab’s Take

Varya pulls off “20x cost difference” AND “cultural recognition”—that’s pretty much unheard of in the AI video generation space. Let’s break down its tech path and business logic.

The key takeaway isn’t the product itself—it’s the development strategy: instead of training a model from scratch, they took Alibaba’s open-source Wan 2.2, used knowledge distillation to compress inference steps from 50 down to 4, achieving 10x speed and major cost reduction. This shows something important—localization doesn’t mean rebuilding everything from the ground up. Finding the right open-source foundation plus targeted reinforcement can get you past both tech and cost barriers simultaneously. The Indian government’s “India AI Mission” mechanism—subsidizing compute in exchange for open model release—also gives us an interesting policy framework to watch: using public resources to build local AI infrastructure rather than letting pricing power sit in the hands of a few commercial platforms. For AI builders, this logic is way more practical than “building cool features”—market entry is often about cost structure, not tech itself.

My advice? Head over to Varya’s website and generate a video yourself, then compare the per-second cost to whatever video tool you’re using now—that price gap will recalibrate where you think the cost baseline for AI video generation actually is.

📅 Source Info

🔗 Further Reading