Tesla(US:TSLA) 自动驾驶之心·2026-01-21 00:51Core Viewpoint - Tesla is not a One Model but a combination of nearly 200 small scene models, as analyzed by overseas sources [4][12]. Group 1: Model Architecture - Tesla's HW4 features two model combinations: Node A with 189 neural networks and Node B with 110, sharing 61 networks between them [4]. - Different models are allocated for various scenarios such as factories, highways, city streets, and destination approaches, with independent end-to-end modules deployed for each [5]. - The system architecture is designed in a modular way, allowing parts to operate independently or in a pipeline collaboration [6]. - HW3 and HW4 share a total of 135 neural networks, with HW4 having significantly larger model sizes compared to HW3 [7][8]. Group 2: Model Operation and Performance - The current large model field is adopting similar approaches, introducing an Agent model where input is routed to relevant models for responses [9][10]. - However, Tesla's Full Self-Driving (FSD) is not yet as intelligent as the Agent models, lacking reasoning capabilities typical in large language models (LLMs) [11][12]. - Tesla's operational frequency of 36 Hz indicates the use of smaller models, which is supported by the bandwidth capabilities of HW4 at 448 GB/s compared to HW3's 68 GB/s [14][15]. Group 3: Engineering and Technology - Tesla is characterized as a company that excels in detailed engineering, often perceived as a high-tech firm [19]. - Elon Musk is recognized for leveraging existing technologies to create superior products, emphasizing the engineering aspect of FSD [21][22]. - The smooth operation of Tesla's systems is attributed not only to computational power and models but also to the rewritten vehicle control operating system, which reduces latency significantly [23].