Tesla FSD v14.3 rolls out with MLIR rewrite, 20% faster reaction times

Details

• Tesla began rolling out Full Self-Driving (Supervised) v14.3 to HW4 vehicles via build 2026.2.9.6, with the headline improvement being a complete rewrite of the AI compiler and runtime using MLIR, delivering 20% faster reaction times.
• The update upgrades reinforcement learning training, enhances the neural network vision encoder for rare and low-visibility scenarios, and strengthens 3D geometry and traffic sign understanding.
• Parking improvements include increased decisiveness in spot selection and maneuvering, with a new "P" icon on the map predicting parking locations before the vehicle arrives.
• Enhanced safety responses to emergency vehicles, school buses, right-of-way violators, and improved handling of small animals and unusual roadside objects sourced from Tesla fleet data.
• The system now recovers from temporary degradations like camera or compute hiccups without driver intervention, reducing unnecessary disengagements.
• Upcoming improvements in the pipeline include pothole avoidance, expanded AI reasoning across all driving behaviors, and enhanced driver monitoring with better eye gaze tracking.

Impact

The MLIR compiler rewrite represents a significant technical foundation upgrade that positions Tesla's FSD stack for faster iteration cycles and reduced latency—critical factors in autonomous driving safety. By centralizing on MLIR, an industry-standard compiler infrastructure, Tesla gains access to broader optimization tooling while signaling deeper technical maturity. The 20% reaction-time improvement is measurable safety gain; in autonomous systems, milliseconds matter for collision avoidance. Fleet-learning integration for edge cases demonstrates Tesla leveraging its deployed base as a distributed testing ground, a competitive advantage Waymo and Cruise have pursued but without Tesla's scale. The focus on parking and recovery from transient failures addresses real user friction points, making FSD more operationally reliable for daily use rather than just highway autonomy.