TurboVLA: Robot AI Matches 7B Models, No LLM Needed

A new vision-language-action model called TurboVLA challenges the assumption that large language models are essential for robotics. This model achieves 97.7% success on a manipulation benchmark without using any language model at all. What's interesting is that TurboVLA runs at 32 Hertz on a consumer GPU, using less than 1 gigabyte of memory. This is five times faster than the dominant open-source VLA model, OpenVLA, while using one-sixteenth of its memory. TurboVLA also matches or beats the accuracy of a much larger, 7-billion parameter model. The key difference is that TurboVLA directly maps visual and language inputs to actions, instead of routing them through a large language model. This change reduces the model's parameters from 7 billion to 0.2 billion. The bottom line: this development suggests that current hardware is sufficient for advanced robotics, potentially speeding up research and development.