Guangzhou, May 6, 2026 — XPENG (NYSE: XPEV, HKEX: 9868), a leading China-based high-tech company, previously released the X-World technical report and demonstrated the practical value of this technology in XPENG's autonomous driving. Recently, XPENG once again announced advancements in world model technology, the X-Cache technical report.

X-Cache leverages the continuity of the physical world to identify reusable image regions while ensuring safety, thereby reducing redundant computations. It can be directly applied to world models in a fast and lightweight manner (without requiring retraining), achieving up to 2.7 times faster denoising inference acceleration for world models. This significantly enhances efficiency and reduces resource consumption.

Reductive yet Reliable, Exploiting Physical Continuity for Cross-Segment Feature Reuse
As autonomous driving enters the model-driven era, high-fidelity simulation of the real world has become a cornerstone for the continuous evolution of driving models. While autoregressive video diffusion-based world models offer high-fidelity, multi-view video generation capabilities, their inference cost and latency remain bottlenecks constraining real-time interaction and large-scale deployment.

XPENG employs fewer steps to refine visuals that closely mirror the real world (a technique known as few-step distillation). However, in this context, traditional acceleration methods, which identify similarities between denoising steps to enable skipping, fail to resolve the issue of slow inference.

The core insight behind X-Cache stems from a physical fact: autonomous driving footage is continuous and evolves smoothly. During driving, elements such as the road surface, roadside trees, and distant buildings change little between the previous frame and the next. Consequently, X-Cache partitions the video into temporally continuous "segments" and compares the intermediate feature similarity within the same layer and at the same denoising step across adjacent segments. If the variation is minimal, previously computed intermediate results are directly reused, and the entire layer computation is skipped. This constitutes the cross-segment caching logic of X-Cache.

In essence, rather than relying on the "step" dimension, where redundancy is already eliminated by few-step distillation, X-Cache optimizes along the novel dimension of "continuous generated segments.

Overall architecture of X-Cache

To ensure the accuracy of cross-segment reuse, X-Cache generates a "fingerprint": it incorporates driving actions (e.g., aggressive steering) alongside visual structure to assess whether current road conditions resemble recent ones, enabling more intelligent reuse. Concurrently, X-Cache features a "safety mechanism" that triggers full computation at critical moments of scene transition, such as turning, lane changing, or traffic light switching (KV update frames), to prevent visual corruption caused by error accumulation.

Consequently, X-Cache significantly enhances the inference efficiency of world models without sacrificing generation quality, offering a viable solution for applications requiring high concurrency and high-frequency invocation.

An Intelligent, Plug-and-Play Utility for Lossless World Model Acceleration
X-Cache is a training-free control logic with cache contents refreshed in real time during generation; its overhead remains manageable compared to the parameter count of the model itself.

Unlike solutions that remain confined to the experimental stage, this intelligent utility has been successfully deployed in XPENG's autonomous driving world model, X-World, operating stably across diverse complex scenarios such as urban roads and highways. By enabling cross-segment computation reuse, X-Cache achieves high compute utilization and inference acceleration, while ensuring generation quality and system stability through multiple mechanisms—demonstrating engineering reliability suitable for large-scale deployment.

Visual Comparison on Urban Expressways: Baseline Model vs. X-Cache

Visual Comparison on Turning Scenarios: Baseline Model vs. X-Cache

X-Cache achieves a 71% block skip rate and delivers 2.6–2.7× measured inference speedup, with virtually no loss in visual quality.

As a physics-oriented simulation engine, X-World constructs inferable and interactive virtual environments, serving as the core infrastructure for model training and continuous evolution. Building on this foundation, X-Cache further addresses efficiency and cost challenges in large-scale simulation, endowing high-quality simulation with the engineering capability to be "runnable, fast-running, and cost-controllable." Supported by this architecture, the performance ceiling of XPENG VLA 2.0 is significantly elevated.