← Back to feed

Avataar AI, a Peak XV-backed startup selected by India’s government AI programme, launched Varya on Friday — an open-weight video generation model designed for Indian cultural contexts and priced at roughly $0.005 per second of output. That is a 20x discount to the $0.10 or more per second charged by global leaders Veo, Kling, Luma, and Runway.

The launch was announced at a press event in New Delhi by S. Krishnan, Secretary of India’s Ministry of Electronics and Information Technology, and positions Varya as a milestone in the country’s push for sovereign AI capability across modalities.

The Technical Architecture

Varya is a distilled model, not a ground-up architecture. It starts from Wan 2.2, Alibaba’s open video generation model, and compresses its inference path from 50 denoising steps to 4. The result:

• Inference speed: 45 seconds to generate a 5-second 720p clip on an NVIDIA H200, versus approximately 1,230 seconds for the base Wan 2.2 model — a 27x speedup
• Cost at hosted service: Rs 0.48 per second (~$0.005 USD), versus $0.10+ for comparable global models
• Output quality: Comparable to Wan 2.2 at full steps, per Avataar’s internal benchmarks
• Steps reduced: 50 denoising steps to 4

The model runs on the subsidised compute infrastructure provided by the India AI Mission, a $1.2 billion government initiative that gives 12 selected startups access to national GPU clusters in exchange for publicly releasing their outputs.

Cultural Specificity

Varya was trained on curated data focused on Indian visual and cultural contexts. The stated goal is to accurately generate content depicting Indian regional variation, festivals, food, clothing, architecture, and everyday environments — domains where global video models trained on Western internet data routinely fail or produce stereotyped outputs.

The claim is not just lower cost but relevance. For a country of 1.4 billion people generating an expanding market for e-commerce video, government communications, and educational content in multiple languages, generic AI video that defaults to Western visual idioms is a practical limitation.

Open-Weight Release

Varya’s weights will be released on India’s AI Kosh portal (aikosh.indiaai.gov.in) along with training data, under terms allowing self-hosting and fine-tuning. That makes it the first Indian government-backed AI video model to release weights publicly, and one of the few culturally-specialised open video models anywhere.

The release structure contrasts with the Glasswing-style restricted-access approach being used for frontier models in the US. India’s model is publicly-available-first, with hosting as an optional monetisation layer.

The Sovereign AI Context

Varya is Avataar’s contribution to a broader Indian government strategy. MeitY Secretary Krishnan stated the government plans to release sovereign AI models periodically across different modalities. Three patterns from this launch are worth tracking:

1. Distillation-first development: Rather than training frontier-scale models from scratch, India’s approach leans on open-weight foundations from Chinese labs (here Alibaba’s Wan 2.2) and applies domain-specific distillation — a practical shortcut given GPU constraints
2. Public infrastructure enabling private output: The India AI Mission’s compute subsidies produced a commercially deployable model with open-weight public access, a structure closer to DARPA-style industrial policy than standard academic grants
3. Price as a differentiator: At $0.005/second, Varya is not competing with Veo 3 on photorealism — it is competing on total cost of ownership for high-volume production workloads in markets where $0.10/second is prohibitive

Avataar is also in conversations with Higgsfield and Adobe Firefly for platform integrations. If the open-weight release generates traction, it becomes the foundation for a broader Indian video AI ecosystem — the same playbook China ran with Wan 2.2 itself.