On June 29, Tesla released FSD v14 Lite, which distills HW4-level V14 intelligence onto HW3 hardware, enabling HW3 vehicles to access reinforcement learning and offline models previously exclusive to HW4. The update also introduces parking, unparking, and reversing functions for the first time. This update significantly extends the lifespan of older HW3 vehicles by bringing cutting-edge autonomous driving capabilities to millions of existing owners without hardware upgrades. It demonstrates Tesla's engineering prowess in model distillation and reinforces its commitment to continuous software improvement across hardware generations. FSD v14 Lite improves navigation, lane merging, pedestrian interaction, traffic light handling, and cut-in responses, while reducing unnecessary braking and enhancing steering smoothness and lane centering. The update also adds full-time speed profile customization for driving style personalization.
Background
Model distillation transfers knowledge from a large, powerful model to a smaller one, allowing the smaller model to run on less capable hardware while retaining much of the larger model's performance. Offline reinforcement learning trains agents on pre-collected datasets without real-time interaction, enabling safer and more efficient model updates. HW3 and HW4 refer to Tesla's custom autonomous driving computer hardware generations, with HW4 offering higher compute capacity.