Tesla rolls out Halloween Mode with FSD Supervised v14.3.12

Tesla is pushing software 2026.38.5, pairing a seasonal Halloween Mode with Full Self-Driving (Supervised) v14.3.12 on Hardware 4 cars and a v14.3 Lite branch for Hardware 3 Model 3 and Y vehicles with AMD Ryzen GPUs. Release notes were published around 10 October 2026; trackers report a wider fleet rollout on 13 October.

Halloween Mode

Tesla’s notes describe a Ghost Costume on the car visualisation, Trick or Treat Mode that plays spooky sounds and flickers lights when visitors approach, a scheduled light show and a Halloween photobooth. The feature is listed for newer Model S, 3, X, Y and Cybertruck on AMD hardware.

What FSD v14.3.12 changes

According to the release notes, v14.3.12 includes:

• An upgraded reinforcement-learning stage and vision encoder for rare and low-visibility scenes, 3D geometry and traffic signs.
• A rewritten AI compiler and runtime using MLIR, which Tesla says yields about 20% faster reaction time.
• Smoother lane bias and following behaviour, more decisive parking, better responses to emergency vehicles, school buses and unusual objects in the path.
• Stronger temporary-fault recovery without forcing a disengagement, and more sensitive driver monitoring (eye gaze and eyewear).
• Unified modelling across Actually Smart Summon, FSD and Robotaxi; drivers can log intervention reasons and see intervention-free distance and streaks in the Self-Driving app.

Tesla stresses that FSD (Supervised) is not autonomous driving: a licensed, attentive driver must supervise at all times.

Other features in 2026.38.5

The same package adds Emergency Drive Away (shift to Drive with a Supercharger cable still attached — emergency only, risk of damage), Supercharger site maps with live stall status on some cars, Set Arrival Energy from the mobile app (v4.58.0+), dynamic headlight leveling on some Model 3 and Cybertruck units, and Automatic Collision Evasion on HW4 that can engage FSD if the driver seems inattentive.

Our take

Halloween Mode is seasonal fun; the lasting question is whether the MLIR runtime and RL training make everyday FSD feel more decisive without adding new failure modes. As always, treat Tesla’s percentage claims as company statements until independent testing confirms them.