In the first half of this year, XPENG — one of Australia's leading Chinese EV brands — has released a series of Physical AI technical reports. These include the multi-view generative world model X-World, the inference acceleration engine X-Cache, the predictive world model X-Foresight, and X-Mind.
TuringViT will underpin three core business scenarios — intelligent driving, the smart cockpit, and the IRON humanoid robot — while providing the industry with a reproducible, low-cost path to training SOTA visual models.
Paper Source: TuringViT: Making SOTA Vision Transformers Accessible to All
The results compare zero-shot performance, training data scale and 1K-resolution throughput of vision transformer encoders. Using just 10% of the training data, TuringViT outperforms leading open-source baseline models.
The Vision Encoder: The Core Bottleneck in Scaling Physical AI
If the world model is where Physical AI perceives and reasons about the laws of the environment, the vision encoder is its "perception gateway": every scene understanding, state judgment, and action decision begins with high-quality visual feature extraction.
As VLM and VLA technologies rapidly move into real-world deployment, industry demands on vision encoders are changing fundamentally: high-resolution images, multi-view inputs and continuous video frames are becoming the norm, while on-device deployment imposes strict latency and compute constraints across different scenarios.
Compute cost bottleneck: standard softmax attention scales quadratically with the number of visual tokens, so training and inference costs surge in high-resolution, multi-frame video scenarios — making real-time on-device deployment difficult.
Data efficiency bottleneck: mainstream high-performance ViTs rely on tens of billions of image-text pairs, but web-scale data is noisy and poorly aligned. The marginal return of simply scaling data is diminishing, making it hard for smaller teams to replicate SOTA results.
Scenario adaptation bottleneck: most vision models are pre-trained at fixed resolutions, which conflicts with the dynamic resolutions and varying aspect ratios required by downstream VLMs/VLAs. Post-hoc adaptation adds R&D cost and introduces performance loss.
The industry's common approach — reusing open-source general-purpose ViTs — can no longer meet the combined demands of performance, latency and customisation in real scenarios such as intelligent driving and embodied robots. TuringViT is XPENG's systematic answer to this industry pain point.
Three Core Innovations: Rebuilding the Training Paradigm for Visual Foundation Models
TuringViT makes simultaneous breakthroughs across architecture, data and training, building a complete technical system of "linear-attention dominance + high-quality data curation + native dynamic resolution" — delivering improvements in accuracy, efficiency and deployability.
Turing Linear Attention: Near-Linear Latency Scaling at High Resolution
To address the compute bottleneck of softmax attention over long sequences, TuringViT introduces Turing Linear Attention (TLA) as its core computation unit, built on a "5-layer linear attention + 1-layer standard multi-head attention" configuration.
This design lets linear attention handle the bulk of global context aggregation, reducing computational complexity from quadratic to near-linear. A small number of standard attention layers is inserted periodically to preserve token-level interaction accuracy, paired with sequence-length-aware normalisation and input-dependent gating.
(Left) The backbone is built on repeated Turing blocks, where Turing Linear Attention (TLA) aggregates global context; its sequence-length-aware normalisation and input-dependent output gating stabilise variable-token training while preserving high-frequency local details.
(Right) Configurations of TuringViT-18L and TuringViT-24L, which share the same patch size (16x16) and use 2D RoPE, scaling embedding and head dimensions with block and total layer counts.
TuringViT launches in two configurations: TuringViT-18L contains 3 groups of Turing Blocks (15 TLA layers + 3 MHA layers), focused on efficient deployment at dynamic resolutions; TuringViT-24L offers a larger model scale.
Measured results show that as input resolution increases, TuringViT's latency curve grows far more slowly than a standard softmax ViT, with the efficiency advantage widening at high resolution. At 1536x1536 resolution, the latency advantage over a standard softmax ViT is especially pronounced.
VISTA-Curation: SOTA-Level Results with Just 10% of the Data
Unlike the industry's blunt "scale the data" approach, TuringViT builds the VISTA-Curation multimodal data curation pipeline, achieving an order-of-magnitude improvement in data efficiency by raising the supervisory value of each sample — shifting from "using more data" to "using better data".
For image-text data, the pipeline upgrades quality through a three-step fine-grained filter: first, low-resolution, blurry and low-texture images are removed; second, diverse candidate captions are generated with multiple models and prompts, cross-validated for visual consistency; third, within a unified alignment framework.
For video data, the pipeline applies full-process curation: long videos are first split into continuous short segments with representative frames sampled evenly; segments that are semantically coherent and information-rich are retained through dual semantic-consistency and motion-consistency filtering; finally, local frame-level detail captions are fused.
In the end, using just 0.85B image-text pairs (about 10% of SigLIP2-L's training data scale), TuringViT achieves an average of 83.6% across six zero-shot classification benchmarks, including ImageNet-1K.
Native Dynamic-Resolution Training: Aligning Downstream Needs from the Pre-training Stage
TuringViT adopts a four-stage progressive native dynamic-resolution training paradigm, adapting to the input characteristics of downstream VLMs/VLAs from the very start of pre-training — moving away from the traditional "fixed-resolution pre-training + post-hoc adaptation" model.
The four stages proceed progressively, aligning with real downstream scenarios step by step:
Visual initialisation: mask-image modelling (MIM) distils EVA02-CLIP-E features to learn fundamental visual representations, strengthening the retention of geometric detail and high-frequency information.
Range-constrained dynamic-resolution training: original aspect ratios are preserved with the long edge constrained to 256-512, letting the model adapt to variable-length visual token sequences at a controlled cost.
Native-resolution fine-tuning: resolution constraints are relaxed to preserve original image sizes and aspect ratios, making visual pre-processing fully consistent with downstream VLM usage.
Image-text-video mixed training: curated video data is added, extending static image representations into transformation-aware unified image-video representations to improve transfer to downstream tasks.
Combined with Dinov3-style 2D rotary position encoding (2D RoPE), the model natively supports inputs of different sizes and aspect ratios without extra interpolation or continued adaptation training. This "VLM-native" design aligns visual pre-training with downstream multimodal systems.
A Unified Foundation for All Scenarios: Strengthening the Physical AI Technical Loop
With the release of TuringViT, XPENG's fully self-developed Physical AI foundation technologies are progressively maturing and expanding. These capabilities serve not only intelligent assisted driving but also a wider range of Physical AI scenarios.
Empowering the Second-Generation VLA: Bringing High-Resolution Perception to the Vehicle
In intelligent driving, TuringViT serves as the core vision encoder of the second-generation VLA model, processing high-resolution, multi-frame dynamic road scenes from multiple surround-view cameras and delivering low-latency, high-quality visual tokens to the predictive world model.
Its near-linear latency profile lets high-resolution visual perception run stably within the vehicle's compute constraints. That upgrades narrow-road driving, navigation-free assisted driving, complex-intersection handling and long-tail traffic recognition for battery electric cars — while further unlocking the potential of on-device large models.
Upgrading Cockpit Interaction: Natural Fusion of Vision and Language
For smart cockpit and integrated driving-parking scenarios, TuringViT's VLM-native characteristics enable more efficient alignment between visual features and language models, delivering higher recognition accuracy and supporting visual inputs of varying frames and aspect ratios across future electric car models.
Strengthening the Embodied Foundation: General Visual Capabilities for the IRON Robot
In the IRON humanoid robot's technical system, TuringViT plays the core role of a "visual retina", providing embodied intelligence with fundamental perception capabilities such as fine-grained object recognition, spatial relationship understanding, manipulable-area detection and dynamic environment tracking.
The unified foundation architecture lets the vast visual experience accumulated in intelligent driving transfer quickly to robotic scenarios, sharply reducing cross-scenario R&D cost. Efficient on-device deployment also fits the embedded compute constraints of robots, accelerating applications such as showroom guidance, industrial inspection and commercial services.
An Open Technical Path for the Democratisation of Visual Foundation Models
TuringViT's value extends beyond XPENG's internal applications: it offers the industry a reproducible, transferable SOTA-level methodology for training visual foundation models — the kind of innovation that matters for electric cars in Australia. Its core architecture, data curation pipeline and training paradigm are not tied to any specific hardware.
Looking ahead, XPENG will continue to expand its high-quality image-text-video data scale, deepen research into temporal modelling and embodied vision, and drive deeper integration between visual foundations and VLM/VLA systems — bringing advanced AI into every journey, from long-range electric cars to the smart cockpit.
Appendix: XPENG Physical AI Foundation Model Papers
