Tesla rolls out FSD v14.3

A self-driving update is mostly about milliseconds. When a car camera sees a yellow light or a school bus, the software has to turn pixels into a steering or braking decision fast enough to matter in the real world. (mlir.llvm.org) (notateslaapp.com) That is why Tesla’s newest Full Self-Driving build is not centered on a new dashboard button or a new map screen. The company says version 14.3, shipped inside software build 2026.2.9.6, rewrites the artificial intelligence compiler and runtime and cuts reaction time by 20 percent. (electrek.co) (notateslaapp.com) A compiler is the translator between a model Tesla trains in the lab and the silicon chip that runs inside the car. If that translator is clumsy, the neural network can be smart on paper and still feel slow on the road. (mlir.llvm.org 1) (mlir.llvm.org 2) Tesla says it rebuilt that translator on Multi-Level Intermediate Representation, which is an open-source compiler framework from the LLVM project designed to optimize code across different hardware targets and abstraction levels. In plain English, it is a better set of factory tools for turning one driving model into something the car can execute with less delay. (mlir.llvm.org 1) (mlir.llvm.org 2) The rest of the release notes read like a list of edge cases that make drivers tense up. Tesla says version 14.3 improves behavior around emergency vehicles, school buses, right-of-way violators, small animals, compound traffic lights, curved-road yellow lights, and objects hanging or leaning into the lane. (notateslaapp.com) (teslascope.com) It also changes lower-speed moments that owners notice every day. Tesla says the update reduces unnecessary lane biasing and minor tailgating, makes parking-spot selection more decisive, and adds a parking pin on the map marked with a “P” icon. (notateslaapp.com) (teslascope.com) Tesla is also leaning harder on reinforcement learning, which is the training method where software gets rewarded for better choices the way a dog gets a treat for the right trick. The company says it upgraded that stage of training and used harder examples from the fleet to improve traffic-light handling and proactive safety around small animals. (notateslaapp.com) (teslascope.com) For now, this is still a limited rollout, not a fleet-wide flip of a switch. Tracking sites showed only a handful of pending installations on April 7, 2026, and Tesla-focused reporting said the first wave was going to early-access Hardware 4 cars in the United States. (teslascope.com) (teslaoracle.com) That hardware detail matters because Tesla’s release notes list version 14.3 for Hardware 4 vehicles including Model S, Model 3, Model X, Model Y, and Cybertruck. Owners on older Hardware 3 systems are not in this release path, so the headline improvements are tied to Tesla’s newer onboard computer. (notateslaapp.com) (teslascope.com) The bigger shift is that car ownership now changes after you buy the car. A brake pedal is fixed at the factory, but a parking routine, a traffic-light response, and even the speed of the model-to-chip pipeline can all change in an overnight download. (electrek.co) (notateslaapp.com) Tesla’s own notes still include the same warning they have used for years: Full Self-Driving is supervised, it does not make the vehicle autonomous, and the driver must stay attentive. So the story here is not that Tesla solved self-driving on April 7, 2026; it is that the company is now shipping deeper changes in the software stack, from training rewards to compiler plumbing, as part of an ordinary consumer update. (notateslaapp.com) (digitaltrends.com)