XPeng unveils TuringViT, a visual encoder that beats top models with 10 times less data

XPeng has launched TuringViT, a new visual encoder designed for VLM and VLA systems. It's built to power self-driving cars, smart cockpits, and humanoid robots. TuringViT comes in two versions, 18L and 24L, and it works much faster with high-resolution images—up to 3 times faster than some competitors at 1536x1536 resolution.

The secret sauce is a data system called VISTA-Curation. It cleans and picks only the best images and captions, so the model learns from quality, not just quantity. Using only 0.85 billion image-text pairs (about 10% of what similar models use), TuringViT scored an average 83.6% accuracy on several zero-shot classification tests, beating models trained on 10 billion pairs. It also does well on image-text search tasks.

TuringViT is a key part of XPeng's tech stack: it's the core visual encoder for their next-gen self-driving system, it helps smart cockpits understand visuals better, and it acts like a "visual retina" for their IRON humanoid robot, giving it object recognition and spatial awareness.