Cheaper, faster, and culturally aware, Avataar's video AI is built for India's scale
Avataar AI launched Varya, a low-cost video model tuned for Indian context. It runs faster than Wan 2.2 and is set to be far cheaper than leading rivals.
Intelligence analysis by GPT-5.4 Mini
Avataar AI is betting that India’s video AI market will be won on cost, speed, and cultural fit. Its new model, Varya, is distilled from Alibaba’s Wan 2.2, is aimed at local use cases, and will be offered as an open-weight release through India’s AI Kosh portal.
Avataar made a video-making robot that is cheaper and faster, like turning a huge delivery truck into a small scooter for busy city streets. It is also taught to notice Indian things like festivals and clothes, so it does not get confused as easily.
Analysis
What Avataar launched
Avataar AI, one of the startups selected under India’s AI Mission, has released Varya, a video generation model built around Indian context. The company says it can recognize things like festivals, food, clothing, and architecture, which addresses a common problem in image and video systems that often produce generic or stereotyped results.
How it was built
Varya was not trained from scratch. Avataar started with Wan 2.2, a publicly available video model from Alibaba, and used distillation to compress its abilities into a smaller system tuned for Avataar’s needs. The company says that redesign cut the model from 50 steps to 4, making it about 10 times faster and much cheaper to run.
Avataar gave one concrete example: on an NVIDIA H200 GPU, Varya can generate a 5-second 720p clip in 45 seconds, while Wan 2.2 takes 1,230 seconds.
Why India is the target
The pricing is the sharpest part of the story. Avataar plans to charge ₹0.48, or about half a cent, per second of video on its hosted service. TechCrunch notes that competing tools such as Veo, Kling, Luma, and Runway often charge $0.10 or more per second.
Peak XV managing director Rajan Anandan said video is especially important in India because consumer internet products there are often video-led rather than text-led, and cost is a major barrier to broad adoption.
Distribution and strategy
Varya will be released as an open-weight model on the government’s AI Kosh portal, along with its training data. That means developers can self-host it or modify it for their own use. Avataar also plans enterprise access and says it is open to partnerships with tools such as Higgsfield and Adobe Firefly.
Key points
- Avataar AI launched Varya, a video model tuned for Indian cultural context.
- The company says Varya was distilled from Alibaba’s Wan 2.2 and runs in four steps instead of 50.
- On an NVIDIA H200 GPU, Avataar says Varya can make a 5-second 720p clip in 45 seconds.
- The hosted service is planned at ₹0.48 per second, far below many major video AI rivals.
- Varya will be released as an open-weight model on India’s AI Kosh portal with training data included.
If Varya performs as advertised, it could make video generation affordable enough for students, teachers, creators, small businesses, and public services in India. Its open-weight release could also help local developers adapt the model instead of depending only on expensive hosted tools.
The model’s usefulness will depend on whether its cheaper, faster design still produces high-quality video in real-world use. India’s broader AI gap in compute and data could also limit how quickly this kind of model scales beyond a few well-funded teams.
