Avataar AI launched a video generation model called Varya that produces a 5-second 720p clip in 45 seconds on an NVIDIA H200 GPU, against 1,230 seconds for the model it was distilled from, Alibaba's Wan 2.2. The Peak XV-backed startup is pricing the hosted service at ₹0.48 ($0.005) per second of video, roughly a 20x discount to Veo, Kling, Luma, and Runway, which typically charge $0.10 or more per second. Varya will also be released as an open-weight model on India AI Kosh, the government's centralized repository for publicly available AI models.
The speed gain comes from compression rather than a new base architecture. Avataar started with Wan 2.2, the publicly available video model Alibaba released earlier, and applied distillation to collapse the inference path from 50 steps down to 4. The result generates video about 10 times faster than the parent model while targeting Avataar's specific use cases in e-commerce content.
Avataar is one of 12 startups selected last year for the India AI Mission, a roughly $1.2 billion government program that offers subsidized GPU compute in exchange for releasing the resulting models publicly. The trade — cheap compute now, open weights later — is producing its first wave of usable outputs, and Varya is among the most concrete demonstrations of what the subsidy is meant to buy.
Key facts
- 01Avataar AI's Varya generates a 5-second 720p clip in 45 seconds on an NVIDIA H200, versus 1,230 seconds for Alibaba's Wan 2.2.
- 02Varya runs in 4 inference steps versus 50 for Wan 2.2, producing video roughly 10 times faster.
- 03Hosted Varya costs ₹0.48 ($0.005) per second versus $0.10+ for Veo, Kling, Luma, and Runway — about a 20x gap.
- 04Avataar is one of 12 startups selected for India's $1.2 billion India AI Mission, which subsidizes compute in exchange for public model releases.
- 05Varya ships as an open-weight model on India's AI Kosh portal, including its training data.
The pitch is explicitly about cost rather than capability ceiling. Peak XV managing director Rajan Anandan framed it as a structural problem for India's market: consumer products there skew video-first, but current frontier video models price out the population-scale use cases that would matter most — education, small business, public services.
“India is a video-first market. We see this across every large consumer internet product in India: video wins over text.”— Rajan Anandan, Peak XV managing director
That positioning is consistent with a broader argument from Indian investors that the country's AI opportunity sits in applications and developer tooling rather than in matching U.S. or Chinese frontier labs on raw model scale. Compute scarcity and limited high-quality training data have kept Indian foundation-model output well behind the U.S., Europe, and China. Distilling open models from Alibaba and others, then re-releasing them with local tuning, is a cheaper path to a usable product.
Avataar says it curated training data specifically to capture cultural detail — Indian festivals, food, clothing, and architecture — that global video models tend to render as generic or stereotyped. Outputs from Veo, Runway, and similar systems have repeatedly drawn criticism for cultural defaults that read as Western. A model trained with Indian reference material is a narrower bet, but it is the bet the local market is asking for.
“Current AI video models are too expensive for population-scale use in India. If video AI is going to reach students, teachers, MSMEs, creators, enterprises, and public services, costs have to come down dramatically. Cost is the biggest unlock for AI adoption in India.”— Rajan Anandan, Peak XV managing director
Open-weighting Varya on India AI Kosh, along with its training data, lets developers self-host or fine-tune it without going through Avataar's API. The company is also pitching it to enterprise customers and says it is open to integrations with video tools including Higgsfield and Adobe Firefly. A demo is available now on Avataar's website using text prompts or reference images.
The broader policy backdrop is aggressive. IT minister Ashwini Vaishnaw said earlier this year that India aims to attract $200 billion in AI investment by 2028 and more than double its GPU capacity within six months. The India AI Mission is the supply-side leg of that push: cheap compute to a small cohort of selected startups, with an open-release condition that compounds the public investment into shared infrastructure.
The caveat is that distilled models inherit the limits of their parents. Varya's quality ceiling is bounded by what Wan 2.2 can already do, and Alibaba's model is not the current frontier — Veo and Runway lead on fidelity and motion coherence in independent comparisons. Avataar is making a deliberate trade: a measurable step down in absolute quality in exchange for roughly 20x lower cost and 27x faster generation. Whether that lands depends on the workload. E-commerce product video at population scale is exactly the use case the math favors; high-end creative work is not.
Varya is the clearest data point yet that the India AI Mission's bet — subsidize compute, mandate open release, accept that the models will be distilled rather than novel — produces shippable products quickly. For the global video-generation market, a credible open-weight option priced 20x below the incumbents reshapes the floor of what enterprise buyers will pay per second, particularly in price-sensitive regions outside the U.S. and Europe. Veo, Kling, Luma, and Runway are not directly threatened on quality, but their pricing power in emerging markets just got harder to defend.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
