Xpeng completes the visual foundation of physical AI

```

On July 21, XPeng Group announced the official launch of the TuringViT high-efficiency visual encoder, targeting VLM (Vision-Language Model) and VLA (Vision-Language Action Model) application scenarios, with systematic restructuring in architecture design, data paradigm, and training process. According to XPeng, this encoder will be used in XPeng’s smart driving, smart cockpit, and IRON humanoid robot business scenarios.

In VLM/VLA systems, the visual encoder is commonly regarded as the "perceptual gateway" of physical AI, responsible for converting images or videos into visual features for subsequent language or action modules to process.

As VLM/VLA technology is applied in real-world scenarios, high-resolution images, multi-view inputs, and continuous video frames present higher demands on the efficiency and temporal processing abilities of visual encoders.

XPeng believes that the industry’s prevalent "reuse of open-source universal ViT" approach can hardly meet the requirements of smart driving, embodied robotics, and other scenarios for performance, latency, and customization at the same time.

XPeng summarizes the design of TuringViT in three dimensions.

In terms of architecture, TuringViT is mainly based on Turing Linear Attention (TLA) and retains a small amount of standard multi-head attention, forming a hybrid Turing Block architecture; for high-resolution inputs, computational complexity shifts from quadratic scaling to near-linear.

According to XPeng's published test data, at a resolution of 1536×1536, the inference throughput of TuringViT-18L is 3.04 times that of Seed1.5-ViT.

This encoder offers two versions: TuringViT-18L includes 3 sets of Turing Blocks, aimed at deployment with dynamic resolution; TuringViT-24L includes 4 sets of Turing Blocks, focusing on enhanced representation ability.

On the data side, TuringViT adopts the VISTA-Curation multimodal data governance pipeline. According to XPeng, this pipeline applies multi-stage filtering, re-description, and scoring to image-text and video data, improving the supervisory information of single samples. TuringViT is pre-trained on 0.85 billion image-text pairs, about 10% the data scale of SigLIP2-L. On six zero-shot classification benchmarks including ImageNet-1K, TuringViT-24L's average accuracy is 83.6%.

For training, TuringViT uses a four-stage progressive native dynamic resolution training scheme, adapting to the input characteristics of downstream VLM/VLA during pre-training, and with 2D rotary position encoding, supports inputs of different sizes and aspect ratios. XPeng states that this design reduces reliance on the "fixed resolution pre-training + post-adaptation" path.

According to XPeng, TuringViT will be used in three types of business scenarios.

In intelligent driving, XPeng says TuringViT is the core visual encoder of its second-generation VLA model, tasked with processing high-resolution images from multiple surround-view cameras and multi-frame dynamic road scene inputs, providing visual tokens for predictive world models.

In smart cockpit scenarios, TuringViT's native VLM features are used to align visual features with language models, and support visual inputs of different formats and proportions.

Within the technical framework of the XPeng IRON humanoid robot, XPeng positions TuringViT as the foundational visual module, responsible for object recognition, spatial relationship understanding, operable area detection, and dynamic environment tracking.

At the March 20 earnings conference, He Xiaopeng said the XPeng IRON humanoid robot is scheduled for mass production by the end of 2026, with a monthly capacity target of over 1,000 units, and will be prioritized for commercial use in XPeng stores.

XPeng states that the architecture design and training process of TuringViT do not depend on any specific hardware platform, and can provide the industry with a reproducible training path for large vision models.

From a technology strategy perspective, TuringViT fills in the visual encoder gap in XPeng’s physical AI technology system.

Previously, XPeng has successively revealed multi-view generative world model X-World, world model inference accelerator engine X-Cache, predictive world model X-Foresight, and the X-Mind technology framework. TuringViT is positioned at the perception input end of this physical AI technical system.

However, for visual encoders to scale up from technical launch to real-world deployment, engineering adaptation, computing power deployment, and scenario validation are still required.

Whether TuringViT can operate steadily in three differentiated physical environments—vehicle, cockpit, and robot—remains to be observed. XPeng states it will continue to expand the scale of high-quality image-text and video data, and deepen technical exploration in temporal modeling and embodied vision.

Risk Notice and DisclaimerThe market has risks; investment needs caution. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable to their particular circumstances. Investment based on this information is at your own risk. ```