Avataar AI has launched Varya, a video generation model built from Alibaba’s open Wan 2.2 model and distilled for faster, cheaper output. The company says Varya can generate 5-second 720p clips on an NVIDIA H200 in 45 seconds, versus 1,230 seconds for Wan 2.2. Avataar plans to release the model and training data through India’s AI Kosh portal while offering hosted access at about $0.005 per second.
Latent Space interviews Ethan He, who led Grok Imagine at xAI, about building the product in three months. The episode contrasts video generation with world models and explores why video agent models may become an important next step. It also argues that Grok Imagine remains underrated, while the supplied description does not include architecture details or benchmark results.