Avataar Builds Varya Around India’s Video Economics

Avataar AI has launched Varya, a video model designed for Indian use cases where cost, speed and local context matter as much as raw generation quality.

The company is one of the 12 startups selected for the India AI Mission, a roughly $1.2 billion government initiative that gives chosen startups subsidized GPU compute in exchange for releasing models publicly.

The model is aimed at a market where video dominates consumer internet behavior and where expensive AI video tools can limit adoption by students, teachers, MSMEs, creators, enterprises and public services.

Avataar says Varya is trained to recognize cultural details such as food, clothing, architecture and festivals, areas where image and video models often produce generic results.

A Leaner Model Is The Core Product Claim

Avataar did not build Varya from scratch.

It started from Alibaba’s publicly available Wan 2.2 video generation model and used distillation to create a leaner version for its own use cases.

The company says Varya runs in four steps rather than Wan 2.2’s 50, produces video 10 times faster and costs far less to operate.

The performance claim is specific.

Avataar’s benchmark uses an NVIDIA H200 GPU and puts Varya at 45 seconds for a 5-second 720p clip.

The comparable figure given for Wan 2.2 is 1,230 seconds.

Pricing is another part of the pitch: Avataar’s hosted service is planned at ₹0.48 ($0.005) for each second of video, while models including Veo, Kling, Luma and Runway are described at $0.10 or more per second.

Open Weights Tie The Launch To Policy

Varya is set for release as an open-weight model on AI Kosh, India’s centralized portal for public AI models and datasets.

Avataar also plans to release the training data, allowing developers to self-host or modify the model.

Enterprise access is also planned, and named partnership targets include Higgsfield and Adobe Firefly.

That structure connects a product launch to India’s wider AI policy.

The India AI Mission is trying to offset limits in compute access and quality data availability by supporting selected model builders.

The policy backdrop also includes Ashwini Vaishnaw’s target of $200 billion in AI investment by 2028 and a plan to more than double GPU capacity within six months.

The Test Is Adoption Beyond A Demo

Avataar’s launch is strongest where the evidence is concrete: model release plan, open-weight distribution, GPU benchmark, price per second and cultural-context training claims.

The unresolved issue is customer proof.

The available details do not disclose paying user numbers, enterprise contracts, conversion rates from the public website trial or independent quality benchmarks.

Varya is best read as an India AI infrastructure-and-application signal rather than a routine app update.

It shows how lower inference cost, localized data and public compute support can shape model strategy in a market that may not follow the same economics as U.S., European or Chinese AI video platforms.