Tesla updates Full Self‑Driving v14.3

Tesla just shipped a self-driving update that is mostly invisible to drivers. The headline feature is not a new button or a new screen, but a rewrite of the software layer that turns Tesla’s artificial intelligence models into code the car can run in real time, and Tesla says that rewrite cuts reaction time by about 20 percent. (electrek.co) That matters because modern driver-assistance systems are not one giant program with fixed rules. Tesla’s system relies on neural networks, which are pattern-finding models trained on video and driving data, and those models have to make decisions fast enough to handle a lane change, a yellow light, or a pedestrian stepping near the road. (tesla.com, electrek.co) A compiler is the translator between a model built by engineers and the chips inside the car. If that translator is inefficient, even a strong model can hesitate, waste computing power, or take longer to react when the scene changes. (electrek.co, tesla.com) Tesla says Full Self-Driving v14.3 rewrites that compiler and its runtime from the ground up using Multi-Level Intermediate Representation, a compiler framework known as Multi-Level Intermediate Representation, or MLIR. The company says the result is faster reaction time in the car and faster model iteration for Tesla’s own engineering teams. (electrek.co, notateslaapp.com) The update is rolling out as software build 2026.2.9.6 for Hardware 4 vehicles, including Model S, Model 3, Model X, Model Y, and Cybertruck. Reports on April 7, 2026 described the release as an initial rollout to early public testers after a short internal employee validation period. (electrek.co, notateslaapp.com) Tesla did not frame v14.3 as a jump to unsupervised autonomy. The product name remains Full Self-Driving (Supervised), which means the human driver is still expected to watch the road and take over when needed. (electrek.co) Beyond the compiler rewrite, Tesla says it upgraded the reinforcement learning stage used to train the driving network. Reinforcement learning is a training method that rewards better choices, and Tesla says the new version improves performance across a wide variety of driving scenarios. (electrek.co) Tesla also says it upgraded the vision encoder, which is the part of the system that turns camera images into something the driving model can understand. In v14.3, Tesla says that improves handling in rare and low-visibility scenarios, strengthens three-dimensional geometry understanding, and expands traffic sign understanding. (electrek.co) The release notes show Tesla focusing on edge cases instead of headline stunts. Tesla says v14.3 improves responses to emergency vehicles, school buses, right-of-way violators, small animals, and unusual objects that extend, hang, or lean into the vehicle’s path. (electrek.co) Tesla also says it reduced unnecessary lane biasing and minor tailgating behaviors. Those are the kinds of problems that do not always show up in a demo video but heavily shape whether a ride feels smooth, cautious, and predictable to passengers and nearby drivers. (electrek.co) Parking got attention too. Tesla says v14.3 increases decisiveness in parking spot selection and maneuvering, and adds a parking location pin on the map marked with a “P” icon. (electrek.co, notateslaapp.com) Another notable change is how Tesla says it trains on rare events. Tesla’s public artificial intelligence page says its networks learn from complicated scenarios sourced from millions of vehicles, and the v14.3 notes repeatedly mention “hard” examples and infrequent events pulled from the fleet to improve behavior at complex intersections and around unusual hazards. (tesla.com, electrek.co) That pattern is important because it says a lot about where Tesla is in 2026. The company is still chasing better autonomy through software rewrites, new training methods, and better handling of exceptions, not through a single breakthrough that suddenly removes the need for supervision. (electrek.co, tesla.com) For fleet operators and engineers, v14.3 is a reminder that self-driving progress often comes from shaving delay out of the stack and feeding the model better examples. A 20 percent reaction-time claim, better recovery from temporary system degradations, and stronger handling of rare scenarios all point to the same strategy: make the system a little faster, a little calmer, and a little less brittle one update at a time. (electrek.co, notateslaapp.com)