Varya 14B Distills Wan 2.2 for $0.005/Sec Video Generation

Avataar AI has released Varya, a 14-billion-parameter open-weight video generation model optimized for Indian cultural contexts and low-cost AI inference. By applying heavy distillation techniques to Alibaba’s Wan 2.2 architecture, the Bengaluru-based startup has reduced the cost of video generation to $0.005 per second. The release targets population-scale adoption across micro, small, and medium enterprises (MSMEs), rural education platforms, and government citizen services.

Distillation and Inference Benchmarks

Varya achieves its speed through an aggressive step-reduction distillation process. The baseline Wan 2.2 model requires 50 computational steps to generate a standard video sequence. Varya compresses this workflow down to just 4 steps.

This structural change heavily reduces the necessary compute time on standard hardware. When rendering a 5-second, 720p video on an NVIDIA H200 GPU, Varya completes the task in 45 seconds. The base Wan 2.2 model requires approximately 1,230 seconds for the identical workload.

Metric	Wan 2.2	Varya
Generation Steps	50	4
Inference Time (5s, 720p)	1,230s	45s
Hardware	NVIDIA H200	NVIDIA H200

Cultural Alignment

Global foundational models heavily index on Western visual concepts and aesthetic assumptions. Avataar AI trained Varya on a curated dataset specifically designed for the Indian subcontinent to eliminate the need for complex prompting workarounds. The model accurately recognizes and renders local festivals, regional architectural styles, national cuisine, and traditional attire such as sarees and turbans.

Pricing Economics

The primary constraint on generative video adoption is the unit cost per output. Avataar AI explicitly prices Varya at ₹0.48 per second, which converts to approximately $0.005.

This pricing tier alters the production math for developers building high-volume video applications. Proprietary global platforms like OpenAI’s Veo, Luma, Kling, and Runway typically charge $0.10 or more per second of generated video. Varya operates up to 27 times cheaper than these established alternatives while maintaining domain-specific accuracy for regional content.

IndiaAI Mission and Open Weights

The Indian government heavily subsidized Varya’s development through the $1.2 billion IndiaAI Mission. Avataar AI secured access to national AI compute infrastructure as one of 12 selected projects under the federal initiative, accelerating the distillation and training timeline.

The model will be distributed as an open-weight download on the government’s AI Kosh portal. This distribution method allows developers to self-host the model or modify the weights for highly specific enterprise use cases. Avataar AI is also actively building integrations to surface Varya within existing creative workflows, including Higgsfield and Adobe Firefly. A trial version supporting text-to-video and image-to-video generation pipelines is currently live on the company’s website.

If you are building media applications for the Indian market, downloading the open weights from AI Kosh allows you to establish a predictable, fixed-cost inference pipeline. Test the prompt alignment using the web trial before provisioning dedicated H200 clusters for your production deployment.