Skip to main content

XPeng's TuringViT: Ten Times Less Data, Yet Outperforms SOTA

Visual large models have long faced a daunting barrier: to achieve top-tier performance, you need massive datasets and enormous computing power. This has kept advanced visual capabilities in the hands of only a few leading teams. But on July 21, XPeng Group introduced TuringViT, an efficient visual encoder that promises to change that. The company claims it systematically rethinks the visual encoder from architecture to data to training, offering a reproducible, low-cost path to state-of-the-art visual Transformers.

TuringViT is designed for the era of VLM and VLA, supporting three key business scenarios: intelligent driving, smart cockpits, and IRON humanoid robots. It breaks through on three fronts—architecture, data, and training—building a system centered on linear attention, high-quality data governance, and native dynamic resolution. The result? Threefold improvements in performance, efficiency, and practicality.

Image

Architecture: Efficiency at Scale

TuringViT comes in two flavors. The TuringViT-18L packs three Turing Blocks with 15 TLA layers and 3 MHA layers, focusing on efficient deployment under dynamic resolution. The TuringViT-24L includes four Turing Blocks with 20 TLA layers and 4 MHA layers, boosting representation capabilities while keeping linear computation as the backbone. Test data shows that as input resolution increases, TuringViT's latency grows much slower than standard softmax ViT. At 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. This clears major hurdles for edge deployment of high-resolution perception, multi-camera input, and long-duration video processing.

Data: Quality Over Quantity

While the industry often throws more data at problems, TuringViT takes a smarter approach. It introduces the VISTA-Curation multimodal data governance pipeline, which focuses on enhancing the supervisory value of each sample rather than just scaling up. For image-text data, the pipeline works in three steps: first, it filters out low-resolution, blurry, or low-texture images; then, it generates diverse candidate captions using multiple models and prompts, cross-validating visual consistency; finally, it scores each sample based on relative image-text alignment, text information volume, and ambiguity, selecting high-quality annotations while eliminating weak or ambiguous ones. For video data, the pipeline splits long videos into short segments, samples representative frames, and uses semantic and motion consistency to filter effective clips. It then combines local frame-level captions with global temporal captions to produce video annotations that capture transformations, helping the model learn robust visual representations from continuous scenes.

Image

Results: Less Data, Better Performance

The payoff is clear. TuringViT used only 0.85 billion image-text pairs—about one-tenth of the data used to train SigLIP2-L—yet achieved an average accuracy of 83.6% on six zero-shot classification benchmarks, including ImageNet-1K. That surpasses mainstream open-source baselines trained on 10 billion data points. It also showed significant advantages in image-text retrieval tasks on COCO and Flickr30K. In short, it gets better results with a fraction of the data.

Image

Training: Dynamic from Day One

TuringViT adopts a four-phase progressive native dynamic resolution training paradigm. From the very first day of pre-training, it adapts to the input characteristics of downstream VLM and VLA, moving away from the old approach of fixed-resolution pre-training followed by post-hoc adaptation.

Key Points

  • TuringViT uses only 0.85B image-text pairs (one-tenth of typical datasets) yet outperforms SOTA baselines.
  • Its architecture features linear attention, enabling efficient high-resolution processing.
  • The VISTA-Curation pipeline ensures high-quality data through rigorous filtering and captioning.
  • A four-phase dynamic resolution training paradigm eliminates the need for fixed-resolution pre-training.