VideoChat3 and RoboTTT Challenge Single-Backbone AI with Specialized Architectures

Two new AI systems, VideoChat3 and RoboTTT, push back against the dominant trend of using one shared transformer backbone for vision, language, and action tasks. VideoChat3 employs a dedicated video-centric backbone with just 4 billion parameters, halving visual token counts while outperforming larger open-source multimodal models on video benchmarks. RoboTTT takes a different approach, integrating test-time training into existing vision-language-action policies to extend context length for robotic tasks, boosting manipulation performance by 87 percent. It also successfully completed a five-minute, ten-stage assembly task that no prior baseline model could finish. Together, the two systems raise the question of whether modality-specific architectures could broadly replace monolithic designs, though questions around training costs and untested deployment scenarios remain open.