Affordable, Faster, and Culture-Savvy: Avataar’s Video AI Designed for India’s Massive Market Needs
Table of Contents
You might want to know
How did Avataar AI reduce video generation time and costs so dramatically compared with existing models?
In what ways does the new model address cultural specificity for India, and how will the government support affect adoption?
Main Topic
India has been slower than the U.S., Europe, and China in releasing large-scale AI model outputs, with only a handful of startups publicly sharing models. Much of the early activity has centered on language and voice models, leaving video generation less accessible at scale. To accelerate development and broaden access, the Indian government created the India AI Mission, a roughly $1.2 billion program that includes subsidized GPU compute for selected startups in return for publicly releasing their models. This program aims to lower the barrier to entry for model development and encourage a more active local ecosystem.
One notable beneficiary is Avataar AI, a startup backed by Peak XV that focuses on video tools for e-commerce and other visual use cases. Avataar released a video model named Varya that is built specifically to understand local Indian context — such as identifying festivals, regional clothing, food, and architectural cues. Rather than training a large video model entirely from scratch, Avataar used an existing publicly available foundation model, Alibaba’s Wan 2.2, and applied a distillation technique. Distillation compresses a teacher model's capabilities into a smaller, faster student model tailored for particular tasks and deployment constraints.
The outcome of that process is a leaner model that requires far fewer steps to generate video. Where Wan 2.2 typically takes around 50 steps for generation, Varya operates in about four steps. That reduction translates into substantial speed and cost improvements: on an NVIDIA H200 GPU, Varya can produce a 5-second 720p clip in approximately 45 seconds, compared with roughly 1,230 seconds for Wan 2.2. In practical terms, this makes video generation about 10 times faster for specific target workloads.
The price difference is particularly striking: Avataar plans to charge about ₹0.48 (roughly $0.005) per second of generated video on its hosted service. That rate is approximately 20 times cheaper than many current offerings such as Veo, Kling, Luma, and Runway, which commonly price services at $0.10 or more per second. Lowering cost is central to achieving broad adoption in India, a market where video content is primary across consumer internet products and where population-scale use depends on dramatically reduced prices.
Beyond cost and speed, cultural relevance is a key selling point for Varya. Image and video generation systems trained on generic datasets often miss local nuances, producing stereotyped or culturally tone-deaf outputs. Avataar reports that it curated training data to help Varya recognize local festivals, food, clothing, and regional architectural styles. This targeted tuning aims to produce outputs that resonate with Indian users and reduce the frequency of culturally inaccurate results.
In line with India’s emphasis on openness and developer access, Avataar will release Varya as an open-weight model via the government’s AI Kosh portal, which centralizes publicly available AI models and datasets. The release includes the model weights and training data so developers can self-host or adapt the model to their requirements. Avataar also plans to offer the model to enterprise customers and to pursue partnerships with video tool providers. A public demo is available on Avataar’s website, where users can try text-prompt or reference-image driven generation.
Varya’s launch highlights a pragmatic approach in which India focuses on delivering practical applications and nurturing a developer ecosystem rather than competing directly on building massive foundation models. The slower pace of foundational model development in India has been attributed to limited compute resources and a shortage of high-quality, localized training data. Programs such as the India AI Mission are designed to narrow that gap by providing subsidized compute and incentives for startups to publish models publicly.
The broader policy context is also ambitious: the Indian government and industry leaders have set substantial targets for AI investment and infrastructure expansion. For instance, India has stated aims to attract large-scale AI investment and to significantly expand GPU capacity over short timelines. These efforts may accelerate the creation, deployment, and adoption of locally relevant AI models, particularly in sectors where cost and cultural alignment matter — including education, small business tools, content creation, and public services.
Key Insights Table
| Aspect | Description |
|---|---|
| Key Fact 1 | Avataar used distillation on Alibaba's Wan 2.2 to produce Varya, reducing generation steps from ~50 to ~4. |
| Key Fact 2 | Varya generates a 5-second 720p clip in ~45 seconds on an NVIDIA H200, ~10x faster than Wan 2.2. |
| Key Fact 3 | Planned pricing is ~₹0.48 ($0.005) per second—about 20x cheaper than many competitors. |
| Key Fact 4 | Varya is trained on curated local data to better capture Indian cultural nuances and will be released open-weight on AI Kosh. |
| Key Fact 5 | The India AI Mission provides subsidized compute to selected startups to stimulate model development and public release. |
Afterwards...
Looking ahead, several technological and ecosystem priorities could help India expand its presence in AI-driven media. Improving access to high-performance compute — including more affordable GPUs and cloud credits — will remain important for local model development and experimentation. Equally crucial is building larger, higher-quality, and more diverse datasets that capture regional languages, cultural practices, and visual styles while respecting privacy and copyright constraints.
Advances in model compression and efficient architectures (such as distillation, quantization, and sparsity) will continue to be valuable for bringing multimedia AI to resource-constrained environments and price-sensitive markets. Continued emphasis on open-weight releases and shared datasets can catalyze developer innovation and localized applications across education, MSMEs, government services, and creative industries.
Ultimately, the combination of targeted technical work — making models both efficient and culturally aware — with public policy that increases compute access and incentivizes openness could enable India to lead in applied AI use cases at population scale rather than competing solely on foundational model creation.