Avataar launches Varya, India's low-cost video model

Avataar launched Varya, a 14-billion-parameter distilled AI video model, at a New Delhi event on June 12, 2026 attended by MeitY Secretary S Krishnan, according to News9 and Outlook Business. Avataar says a distillation technique cuts video generation from 50 steps to 4 steps, pushing its internal benchmark cost to about Rs 0.48 per second of video, which the company and both outlets describe as up to 10 times more cost-efficient than several leading global video models (other coverage cites larger multiples against specific rivals like Alibaba's Wan 2.2). Model weights have reportedly been published to the public AI Kosh portal under the IndiaAI Mission, and Avataar says a technical report on Varya's architecture and benchmarks is coming. Coverage frames the launch around localized cultural content and lower-cost access for creators, educators, and small businesses.
Varya's headline claim, cutting a 50-step video-generation pipeline to 4 steps via distillation, is a concrete data point in the broader industry shift toward smaller, cheaper generative models; if Avataar's own cost benchmark holds up, it puts sub-rupee-per-second video generation within reach of price-sensitive users who could never justify the economics of Veo- or Wan-class models. The more consequential move for practitioners may be that Avataar says it is publishing model weights to India's public AI Kosh portal, a fully open step most video-model vendors don't take.
What happened
According to News9 and Outlook Business, Avataar launched Varya, a 14-billion-parameter video generation model, at a press event in New Delhi on June 12, 2026, attended by S Krishnan, Secretary of the Ministry of Electronics and Information Technology. Krishnan called it "the kind of research-led capability building that we seek to enable," while Avataar CEO and co-founder Sravanth Aluru said "affordability is not a feature, it is a prerequisite" for a market of 1.4 billion people. Avataar built Varya with support from the IndiaAI Mission, including access to subsidized national compute. The company's internal benchmarks show Varya generating video at about Rs 0.48 per second, which Outlook Business and News9 both describe as up to 10 times more cost-efficient than several leading global video models; other outlets have cited larger multiples for specific comparisons (Inc42 reported roughly 27 times cheaper than Alibaba's Wan 2.2 using a rounded Rs 0.50-per-second figure). Avataar says it will publish a full technical report on the model's architecture, distillation method, and benchmarks.
Technical context
Model distillation compresses a larger "teacher" model's behavior into a smaller, faster "student" model. Avataar and multiple outlets describe Varya's version of this as reducing video generation from 50 inference steps to 4. If the reported speed and cost figures hold up under independent testing, that is a meaningful compute reduction, though no outlet has yet published independent benchmarks of Varya's output quality, temporal consistency, or how it compares on standard video-generation evaluation sets. Moolpost additionally reports Varya was trained on roughly 40,000 India-specific cultural datasets and uses techniques including classifier-free guidance and distribution-matching distillation, and that model weights have been published to the public AI Kosh portal; these more specific technical claims come from a single lower-tier outlet and have not been independently corroborated.
What to watch
Worth tracking: Avataar's promised technical report and whether it includes reproducible benchmarks; independent third-party evaluation of Varya's visual fidelity and temporal coherence against Veo, Wan 2.2, and other rivals; and adoption signals from the education, MSME, and public-communication use cases the company is targeting. Because the cost and quality claims currently rest on the company's own figures as relayed by Indian trade press, practitioners should treat the per-second pricing as a vendor claim until benchmarked independently.
Key Points
- 1Avataar's Varya cuts video-generation inference from 50 steps to 4 through model distillation, per the company's own technical description.
- 2Model weights have reportedly been published to India's public AI Kosh portal, an unusually open release among video-generation vendors.
- 3Cost and quality claims rest mainly on Avataar's own benchmarks so far; independent verification and a promised technical report remain outstanding.
Scoring Rationale
Well-corroborated regional model launch: two independent quality outlets (News9, Outlook Business) confirm the same verbatim executive and government quotes, 14B parameter scale, and the 50-to-4-step distillation claim, and the open-weights AI Kosh release is a genuinely notable practitioner signal. Kept below the notable/major threshold because all cost and quality figures are still vendor-supplied with no independent benchmarks.
Sources
Primary source and supporting public references used for this report.
View 4 more sources
- Varya AI model launched in India, can make video at less than 50 paisa per secondnews9live.com
- Avataar Launches Varya, India's First Distilled Video AI Model Under IndiaAI Missionoutlookbusiness.com
- Meet Varya, India's own video AI model that generates videos at just Rs 0.48 per secondmsn.com
- Avataar AI launches Varya AI with Shravant Alurumoolpost.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
