Tesla rolls FSD v14.3 OTA

A Tesla that is already moving itself through city streets still spends most of its time doing one basic job: turning camera video into a decision before the next second arrives. Tesla says Full Self-Driving v14.3 speeds up that handoff by about 20% with a rewritten artificial-intelligence compiler and runtime pushed over the air this week. (notateslaapp.com) A compiler is the software layer that turns a model’s instructions into something the car’s computer can actually run, like rewriting a recipe so a specific kitchen can cook it faster. Tesla’s release notes say v14.3 was rebuilt “from the ground up with MLIR,” a machine-learning compiler framework, and that the payoff is quicker reaction time plus faster model iteration. (notateslaapp.com) The system this update feeds is still not a robot driver you can ignore. Tesla’s own support page says Full Self-Driving (Supervised) can make lane changes, turns, and parking moves, but it “does not make your vehicle fully autonomous” and still requires an attentive human ready to take over. (tesla.com) Version 14.3 also changes what the car sees when the world gets messy. Tesla says it upgraded the vision encoder, which is the part of the neural network that turns raw camera frames into a map of objects and space, to work better in rare and low-visibility scenarios and to improve three-dimensional geometry understanding. (notateslaapp.com) That shows up in very specific edge cases instead of a vague promise of “better driving.” Tesla’s release notes name compound traffic lights, curved-road signals, yellow-light stopping, objects hanging into the lane, small animals, school buses, emergency vehicles, and right-of-way violators as targets of the new training. (notateslaapp.com) The training change is just as important as the code change. Tesla says v14.3 upgraded the reinforcement-learning stage of training and now pulls “hard” examples and infrequent events from the wider Tesla fleet, so one awkward intersection or strange obstacle seen by one car can help train many others later. (notateslaapp.com) Parking got its own pass too, which matters because low-speed parking lots are full of odd angles, pedestrians, and unclear lane markings. Tesla says v14.3 makes parking-spot selection more decisive and adds a map pin with a “P” icon so the car can show the location it predicts for parking. (notateslaapp.com) Another line in the notes hints at a quieter change drivers may notice more often than any headline feature. Tesla says the software is better at handling temporary system degradations and can recover automatically without driver intervention, which means fewer unnecessary disengagements when the system briefly loses confidence. (notateslaapp.com) This rollout is also landing inside a bigger Tesla pattern: frequent, layered software pushes instead of one giant annual drop. Tesla’s v14 trial page shows the company has already been using Full Self-Driving version 14 as a public trial hook, and user tracking sites are now also seeing a separate 2026.8.6 build appear alongside the v14.3 wave, which suggests Tesla is shipping feature work and broader vehicle software on parallel tracks. (tesla.com) (notateslaapp.com) So the real story in v14.3 is not one flashy new button. It is Tesla trying to shave time off the gap between camera input, model processing, and steering or braking output, while using fleet data to train on the rare moments that usually make supervised driving systems look uncertain. (tesla.com) (notateslaapp.com)