This article is a deep-dive from JudyAI Lab — an AI engineering playbook series with 100+ published guides, 5,000+ weekly readers across 60+ countries, focused on the practical side of running AI agents, trading systems, and content pipelines in production.

📰 Key Takeaways

Indian AI startup Avataar AI has launched a video generation model called Varya, built specifically for Indian cultural context — recognizing local festivals, clothing, food, and architecture. Avataar AI is one of twelve startups selected under the Indian government’s “India AI Mission,” a roughly $1.2 billion program that trades subsidized GPU compute for open model releases.

Varya wasn’t trained from scratch — it’s built on top of Alibaba’s open-source Wan 2.2 video generation model, compressed through knowledge distillation to cut inference steps from the original 50 down to just 4, giving a 10x speed boost and a major cost drop. The numbers are striking: on an NVIDIA H200 GPU, Varya generates a 5-second 720p video in just 45 seconds, compared to 1,230 seconds for the original Wan 2.2.

Pricing is what really stands out about Varya. Avataar AI plans to offer its hosted service at ₹0.48 per second of video (about $0.005), compared to mainstream models like Veo, Kling, Luma, and Runway, which typically charge $0.10+ per second — a gap of roughly 20x. Rajan Anandan, Managing Director at Peak XV, pointed out that India is a video-first market, but existing AI video models are too expensive for mass adoption — costs need to come down significantly to reach students, teachers, small businesses, and public services.

Varya will be released as an open-weight model on the Indian government’s AI Kosh platform, along with its training data, so developers can deploy or modify it themselves. Anyone can try it out right now directly on the company’s website using text or a reference image.

💬 JudyAI Lab Take

Varya pulls off both a “20x cost gap” and “local cultural awareness” at the same time — something that’s basically unprecedented in the AI video generation market, and it’s worth unpacking the technical path and business logic behind it.

The most important lesson from Varya isn’t the product itself — it’s the development strategy. Instead of training a new model from scratch, they built on top of Alibaba’s open-source Wan 2.2, using knowledge distillation to compress inference steps from 50 down to 4, getting a 10x speed boost and a major cost reduction. That says something important: localization doesn’t mean rebuilding everything from zero. Pick the right open-source foundation, add targeted reinforcement, and you can clear both the technical and cost barriers at once. The Indian government’s “India AI Mission” mechanism — trading subsidized compute for open models — also offers an observable policy framework: using public resources to build local AI infrastructure, instead of letting pricing power concentrate in a handful of commercial platforms. For AI builders, this logic is far more practical than “build a cool feature” — the real key to market entry is often cost structure, not the technology itself.

Worth going to Varya’s website and generating a video yourself, then comparing it against the per-second cost of whatever video tool you’re currently using — that gap will make you rethink where the real cost baseline for AI video generation actually sits.

📅 Original Source Info

🔗 Further Reading

References