Tesla rolls FSD Supervised v14.3

Tesla rolls FSD Supervised v14.3 A car driving itself has to do three jobs in a few tenths of a second: see the road, guess what happens next, and choose one action before the moment is gone. If any one of those steps lags, the whole system feels hesitant, like a driver who spots a problem half a beat too late. (notateslaapp.com) Tesla’s Full Self-Driving Supervised system is built around neural networks, which are software models trained on huge numbers of driving examples instead of hand-written rules for every turn, stop, and merge. That approach works best when the system has seen enough messy real-world footage to recognize patterns that humans handle without thinking. (notateslaapp.com) One part of that training is reinforcement learning, which is a way of teaching software by rewarding better choices and penalizing worse ones. It is closer to training a dog with feedback than writing a checklist, because the system learns which actions lead to smoother and safer outcomes over time. (notateslaapp.com) Another part is the vision encoder, which is the section of the model that turns raw camera images into something the rest of the system can reason about. You can think of it as the car’s first draft of the world: lane lines, traffic lights, signs, curbs, animals, and the shape of vehicles ahead. (notateslaapp.com) That first draft gets harder when visibility is poor, when an intersection is oddly shaped, or when an object is rare enough that the system has not seen many examples before. A school bus stopped at an angle, a branch hanging into the lane, or a small animal near the road can all be harder for software than an ordinary sedan in daylight. (notateslaapp.com) There is also a less visible bottleneck underneath the model itself: the compiler and runtime that turn trained software into instructions the car’s onboard computer can execute quickly. Even if the model is smart, slow plumbing underneath it can delay the moment when a good decision becomes a steering or braking command. (electrek.co) That is the backdrop for Tesla’s latest release. On April 7, 2026, Tesla began rolling out Full Self-Driving Supervised v14.3 as software build 2026.2.9.6 for Hardware 4 vehicles in the United States, covering Model S, Model 3, Model X, Model Y, Cybertruck, and not older Hardware 3 cars. (notateslaapp.com) The headline change is a ground-up rewrite of Tesla’s artificial intelligence compiler and runtime using Multi-Level Intermediate Representation, usually shortened to MLIR. Tesla says that rewrite cuts reaction time by 20 percent and also speeds up how quickly its engineers can iterate on future models. (notateslaapp.com) Tesla also says it upgraded the reinforcement learning stage of training, which it credits with broader improvements across many driving situations. In the same release notes, the company says it focused that training on harder examples from the fleet, including small animals, yellow-light decisions, and unusual objects extending into the path of the car. (notateslaapp.com) The vision side got an upgrade too. Tesla says the new vision encoder improves understanding in rare and low-visibility scenarios, strengthens three-dimensional geometry understanding, and expands traffic sign understanding, which points to better performance in fog, glare, awkward intersections, and scenes with confusing depth cues. (notateslaapp.com) Some of the most concrete changes are behavioral. Tesla says v14.3 reduces unnecessary lane biasing and minor tailgating, makes parking spot selection more decisive, adds a predicted parking pin on the map with a “P” icon, and improves responses to emergency vehicles, school buses, right-of-way violators, and temporary system degradations that previously caused extra disengagements. (notateslaapp.com) Tesla’s release notes also list three items that are not in this build yet: broader reasoning beyond destination handling, pothole avoidance, and a more sensitive driver monitoring system with better eye-gaze tracking and eyewear handling. That matters because Tesla is presenting v14.3 not as a finished endpoint, but as a faster base layer for the next round of Full Self-Driving Supervised updates. (notateslaapp.com) The company’s own wording still draws a hard line around what the software is. Tesla says Full Self-Driving is “Supervised,” says the driver must remain attentive, and says the feature “does not make your vehicle autonomous,” which means the technical gains in v14.3 arrive inside a system that still legally and practically depends on a human in the seat. (notateslaapp.com)