Core Insights - The second-generation VLA from XPeng Motors represents a significant shift in the traditional "vision-language-action" framework, enabling direct output from visual signals to action commands without language translation [1] - This model is the first mass-produced physical world model from XPeng, serving as both an action generation model and a physical world understanding model, applicable across various domains including AI vehicles, humanoid robots, and flying cars [1][2] - The second-generation VLA is equipped with a parameter scale in the billions, significantly surpassing the industry standard of tens of millions, and is trained on nearly 100 million clips, equivalent to the driving experience of a human driver over 65,000 years [1] Technical Advancements - The second-generation VLA has achieved a breakthrough in computing power and model architecture, leading to a significant evolution in XPeng's intelligent driving capabilities [1] - The introduction of "Xiaolu NGP" has increased the average takeover mileage on complex roads by 13 times, showcasing the model's generalized learning and intelligent emergence capabilities [1] - The industry-first "navigation-free automatic assisted driving" feature, Super LCC+ human-machine co-driving, can be activated globally without reliance on navigation [1] Future Plans - XPeng's second-generation VLA is set to launch a pioneer co-creation experience in December 2025, with full rollout in the first quarter of 2026 alongside the Ultra model [2] - Volkswagen has been announced as the strategic partner for the launch of the second-generation VLA, and XPeng's Turing AI chip has been designated for Volkswagen [2]

Core Insights - The release of the second-generation VLA by XPeng Motors represents a significant advancement in the automotive industry, introducing a new paradigm in physical modeling that allows for direct output from visual signals to action commands without language translation [1][2] - The second-generation VLA is XPeng's first mass-produced physical world model, capable of driving AI applications across various domains, including AI cars, humanoid robots, and flying cars [1] - The model boasts a parameter scale in the billions, significantly surpassing the industry standard of tens of millions, and has been trained on nearly 100 million clips, equivalent to the driving experience of a human driver over 65,000 years [1] Technological Advancements - The second-generation VLA is built on a fully optimized "chip-operator-model" chain, achieving 2,250 TOPS in the Ultra version, which enhances the capabilities of XPeng's intelligent driving systems [1] - The introduction of "Xiaolu NGP" has increased the average takeover mileage on complex roads by 13 times, showcasing the model's advanced learning and generalization capabilities [1] - The industry-first "No Navigation Automatic Assisted Driving" feature, Super LCC+ Human-Machine Co-Driving, can be activated globally without reliance on navigation, demonstrating the model's emergent intelligence [1] Future Plans - XPeng plans to launch the second-generation VLA for pioneer co-creation experiences in December 2025, with a full rollout in the first quarter of 2026 alongside the Ultra model [2] - Volkswagen has been announced as the strategic partner for the launch of the second-generation VLA, with XPeng's Turing AI chip already secured for Volkswagen [2]

他指出，这是小鹏创新于时代的VLA大模型理解，打破了行业标准去思考"能不能拆掉Language"，去掉 语言转译，以视觉为核心，追求"百文不如一见"的模型效果。在投入了 3 万卡算力，烧了 20 多亿的研 发费用后，无语言转译方案 VLA迎来"质变"，第二代VLA涌现了全新可能。 一方面，可以直接使用近1亿真实视频数据训练，相当于驾驶65000年遇到的极限驾驶场景总和；另一方 面，可以通过视觉推理进而预测未来场景，也可以基于真实世界仿真生成训练场景。 新浪科技讯 11月5日下午消息，今日，2025小鹏科技日活动上，小鹏汽车CEO何小鹏表示，小鹏第二代 VLA提前开始'物理世界模型'新范式，是继端到端、标准VLA之后，智能驾驶最大的一次飞跃。 责任编辑：何俊熹 他指出，通过3万卡自动驾驶云端智算集群+云端720亿参数基座大模型，模型每五天全链路迭代一次。 芯片-算子-模型的全链路优化和创新，突破行业车端模型能力上限。在Ultra车型上，实现行业车端最高 有效算力 2250 TOPS；针对性优化算子，模型推理效率提升12倍；基于软硬件协同优化，针对性开发模 型编译器和软件栈，模型参数较行业主流提升10倍。（罗宁） ...