Visual large models have long been held back by an invisible barrier: to achieve top-tier performance, one must first accumulate massive data and computing power, making advanced visual capabilities a luxury of a few leading teams. On July 21, XPeng Group launched TuringViT, an efficient visual encoder, claiming that it systematically restructured the visual encoder from architecture, data paradigms, to training processes, providing the industry with a reproducible, low-cost SOTA-level visual Transformer training path, bringing advanced visual large models from exclusive use by top teams to widespread accessibility across the industry.

The positioning of TuringViT is clear: facing the era of VLM and VLA, it fully supports three business scenarios—intelligent driving, smart cockpit, and IRON humanoid robots. It breaks through in three dimensions—architecture, data, and training—building a technical system dominated by linear attention, high-quality data governance, and native dynamic resolution, achieving threefold improvements in performance, efficiency, and practicality.

image.png

In terms of architecture, TuringViT offers two specifications. TuringViT-18L includes three Turing Blocks, totaling 15 TLA layers plus 3 MHA layers, focusing on efficient deployment under dynamic resolution; TuringViT-24L includes four Turing Blocks, totaling 20 TLA layers plus 4 MHA layers, further enhancing representation capabilities while maintaining linear computation as the mainstay. Test data shows that as input resolution increases, TuringViT's latency growth curve is much flatter than that of standard softmax ViT, with more significant efficiency advantages at high resolutions. At a 1536x1536 resolution, the inference throughput of TuringViT-18L reaches 3.04 times that of Seed1.5-ViT and 2.16 times that of SigLIP2-ViT-L, clearing the core obstacles for edge deployment of high-resolution perception, multi-camera input, and long-time video processing.

Different from the industry's crude approach of increasing data scale, the true sharpness of TuringViT lies in data governance. It has developed the VISTA-Curation multimodal data governance pipeline, no longer pursuing more data but instead achieving a level-up in data efficiency by enhancing the supervisory value of each sample. For image-text data, the pipeline finely screens in three steps: first filtering out low-resolution, blurry, or low-texture poor-quality images; then generating diverse candidate captions using multiple models and prompts and cross-validating visual consistency; finally, scoring comprehensively based on relative image-text alignment, text information volume, and ambiguity in a unified comparison pool, selecting high-visual-fitting, high-semantic-density annotations and eliminating ambiguous, generalizable, and weakly aligned samples. For video data, the pipeline first splits long videos into continuous short segments and uniformly samples representative frames, then uses semantic and motion consistency to filter and retain effective segments. Finally, it combines local frame-level detail captions with global temporal semantic captions to generate video annotations with transformation perception ability, enabling the model to learn more robust visual representations from continuous scenes.

image.png

The improvement in data efficiency directly reflects in the results. TuringViT used only 0.85B image-text pairs, about one-tenth of the data scale used to train SigLIP2-L, yet achieved an average accuracy of 83.6% on six zero-shot classification benchmarks such as ImageNet-1K, surpassing mainstream open-source baselines trained on 10B data. It also showed significant advantages in image-text retrieval tasks on COCO and Flickr30K, truly achieving breakthroughs by obtaining better results with one-tenth of the data.

image.png

In the training process, TuringViT adopts a four-phase progressive native dynamic resolution training paradigm, adapting to the input characteristics of downstream VLM and VLA from the first day of pre-training, completely moving away from the traditional mode of fixed-resolution pre-training followed by post-hoc adaptation.