Tesla Releases FSD 14.3 with Performance Improvements

Tesla rolled out Full Self-Driving v14.3 (build 2026.2.9.6) to Early Access Program members on April 7, 2026, delivering the most technically significant FSD update in the 14.x generation . The headline change is a ground-up rewrite of Tesla's AI compiler using MLIR (Multi-Level Intermediate Representation) architecture, achieving a 20% improvement in reaction time — the largest single-release inference latency gain FSD has seen to date . The update also introduces fleet learning, where Tesla's global fleet contributes rare-event examples (complex intersections, small animal encounters, low-visibility scenarios) back into neural network training .

Behavioral improvements in v14.3 are tangible for daily drivers: more decisive parking spot selection with a visual map pin, better emergency vehicle and school bus response, and fixes to the yellow-light over-braking and stop-sign double-stopping regressions introduced in v14.2 . Musk stated a broader rollout to all HW4 vehicles (Model S, 3, X, Y, and Cybertruck) is set for mid-week, with HW3 owners targeted for a separate FSD 14 Lite branch in mid-2026 . Current FSD subscriptions stand at approximately 1.1 million at $99/month, representing a ~$1.3 billion annualized revenue run rate.

TSLA shares rose +1.49% in premarket trading on the announcement, though the stock remains under pressure — down roughly 11% over the past six months. The MLIR rewrite is strategically significant beyond the headline speed gain: faster inference on existing HW4 chips creates headroom for more capable neural networks without requiring hardware upgrades, directly supporting Tesla's Unsupervised FSD launch targets in Texas and Nevada later in 2026. Bulls cite a long-term robotaxi revenue opportunity reaching $250 billion by 2035; bears note navigation and routing conflicts remain unresolved, and the 20% gain reflects compiler infrastructure improvement rather than a leap in autonomy capability.