Varya: India’s First Distilled Video AI Model 2026 | IndiaAI Mission

Varya is a 14-billion-parameter foundation model that generates video from two types of input:

• Text-to-Video: Enter a text prompt → Varya generates a video clip
• Image-to-Video: Upload an image → Varya animates it into a video

It was not built from scratch. It was distilled from Wan 2.2, a publicly available video-generation model released by Alibaba. Distillation is a machine-learning technique that compresses the capabilities of a large model into a leaner, faster version optimised for specific use cases.

The key technical breakthrough is the reduction in denoising steps. Conventional diffusion-based video models typically require around 50 steps to generate a video. Varya does it in just 4 steps — while maintaining comparable output quality. This is what drives its speed and low cost.

Avataar claims Varya delivers quality comparable to leading global models such as Alibaba’s Wan 2.2 and Google’s Gemini Veo 3. Users follow an “Idea → Video → Story” workflow: enter a prompt or image, generate an initial clip, then extend it with further prompts to build longer narratives.

Cost is Varya’s headline advantage:

• Generates video at roughly ₹0.48 per second (about 50 paise per second)
• About 211 seconds of video per ₹100 spent
• Up to 10 times more cost-efficient than leading global models
• Global competitors (Veo, Kling, Luma, Runway) typically charge more than USD 0.10 per second

Cultural specificity is the second key differentiator. Varya is trained on Indian contexts — regional festivals, attire, food habits, languages, streetscapes, facial features and everyday life. This addresses a recurring weakness in global tools, which often produce generic or stereotyped depictions of Indian settings. Varya is designed for India’s diverse regional realities rather than a single generalised idea of the country.

Varya is a direct product of the IndiaAI Mission, approved by the Union Cabinet in March 2024 with an outlay of about ₹10,372 crore (~USD 1.25 billion) over five years. The mission is led by MeitY and structured around seven pillars:

• Shared compute capacity
• Datasets
• Foundation model development
• Application development
• Skilling
• Start-up financing
• Safe and trusted AI

Avataar was selected among a group of start-ups to build indigenous foundation AI capabilities — alongside language-focused firms Sarvam AI, Soket AI, Gnani AI and Gan AI. The research behind Varya was accelerated by access to subsidised national AI compute infrastructure. India’s shared GPU pool has expanded past 34,000 GPUs, offered to start-ups and academia at below-commercial rates, with a government target of 1,00,000 GPUs by end of 2026.

Varya’s low cost is intended to democratise advanced video generation for segments global tools have priced out:

• Education: Teachers preparing visual lessons in rural classrooms; students creating content
• MSMEs: Affordable region-specific advertisements and promotional content
• Governance: Citizen services, public communication and digital storytelling
• Creators: Independent content creators across urban and rural India
• Enterprise: Avataar plans to make Varya available to enterprise customers and video platforms

The company also intends to publish a technical report detailing the model’s architecture, distillation methodology and performance benchmarks — a step toward open scientific accountability.