Varya
India's first distilled AI video model, trained on local cultural contexts with claimed 10-27x faster generation at a fraction of global competitors' cost.
Overview
Varya is a text-to-video AI model built by Avataar.ai under India's IndiaAI Mission, launched on June 12, 2026. It's positioned as the country's first homegrown video generation model, and its headline claim is dramatic: 10-27x faster inference than models like Sora and Runway, at costs that start at ₹0.48 per second of video (roughly half a US cent). The model uses a distilled architecture, meaning it's derived from a larger foundation model but compressed for speed and efficiency — a legitimate engineering approach that trades some raw capability for dramatically lower latency and cost.
What makes Varya genuinely different from the global video AI crowd is its training data focus. The model was specifically trained on Indian cultural contexts — festivals like Diwali and Holi, regional clothing and textiles, local food, architecture, and social settings. If you've tried generating a Bollywood-style scene or an Indian wedding video in Runway or Pika, you know the results tend to default to Western aesthetics. Varya is built to handle these contexts natively, which gives it a clear niche advantage for Indian content creators, brands, and agencies.
The model is accessible via API with pay-per-second pricing, making it viable for startups and small teams who can't justify the monthly subscriptions of Western platforms. However, it's still very new — launched days ago — and independent benchmarks or third-party reviews are essentially nonexistent. The speed and cost claims are from the team itself, and the model's output quality across diverse prompts remains to be validated by the broader community. For Indian-market content, it's worth watching closely; for global production work, the established players still have deeper feature sets and proven track records.
Key features
Text-to-Video Generation
Generate video clips from text prompts with the model handling scene composition, motion, and visual coherence. Optimized for fast inference through a distilled model architecture.
Indian Cultural Context Training
Trained specifically on Indian cultural data — festivals, traditional clothing, regional food, architecture, and social contexts that global models typically misrepresent or default to Western aesthetics.
Distilled Model Architecture
Uses knowledge distillation to compress a larger foundation model into a faster, lighter version. This enables the claimed 10-27x speed advantage over full-scale models like Sora, at significantly lower compute cost.
Pay-per-Second API
API access priced at ₹0.48 per second of generated video (~$0.006 USD), making it one of the cheapest video generation APIs available. No monthly subscription required.
Pricing
Free tier: Check website for current free tier availability
| Plan | Price | What's included |
|---|---|---|
| API Access | ₹0.48/sec (~$0.006/sec) | Pay-per-second video generation via API; no monthly commitment; optimized for high-volume and cost-sensitive workflows |
Pay-per-second video generation via API; no monthly commitment; optimized for high-volume and cost-sensitive workflows
Pros & cons
Pros
- ✓Extremely low cost — roughly $0.006 per second of video, undercutting every major competitor
- ✓Trained on Indian cultural contexts that global models handle poorly (festivals, clothing, food, architecture)
- ✓Distilled architecture delivers claimed 10-27x speed advantage over full-scale models
- ✓Backed by IndiaAI Mission, signaling government support and long-term investment
Cons
- ×Brand new (launched June 12, 2026) with no independent benchmarks or third-party reviews yet
- ×Output quality relative to Runway, Kling, or Veo 3 is unproven outside the team's own demos
- ×Feature set appears basic compared to competitors offering camera controls, image-to-video, and style transfer
- ×Cultural niche focus means it may underperform on non-Indian visual contexts
