Tesla speeds out FSD v14.2.2.1 in holiday weekend rollout

Alfred Glover

5 hours ago

Tesla’s AI team pushed a fresh Full Self-Driving (Supervised) build, v14.2.2.1, to customers on December 23, just a day after v14.2.2 began rolling out. The new version shares the same feature set as v14.2.2 but appears to focus on stability and bug fixes based on early feedback.

Early driving feedback on v14.2.2.1

Longtime Tesla owner and FSD tester Zack reported detailed impressions after several drives on v14.2.2.1 around rainy Los Angeles. He drove through standing water and areas where lane lines were faded, and he noted that steering kept a steady line with no hesitation or “stutter” in turns. Lane changes were described as confident and smooth, which reminded him of Tesla’s driverless robotaxis running in Austin.

Parking behavior stood out as a clear improvement. The car usually placed itself neatly in a spot on the first attempt, without jittery steering corrections. In one case, the vehicle parked slightly off-center only because a neighboring car had crossed the line, and FSD shifted a few inches to accommodate it. In heavy rain, he said FSD kept track of lanes and turn markings better than a typical driver could, even when paint on the road was hard to see, and it entered new streets in a clean position.

He also took the car on a dark, wet, twisty canyon route and came away impressed. He wrote that it stayed centered, held speed well, and gave a “confidence inspiring” steering feel, adding that it handled the curves better than most human drivers. Other night-drive testers on earlier v14 builds have shared similar views about improved lane discipline and reaction time after dark.

What FSD v14.2.2 brought to the fleet

Version 14.2.2 laid the groundwork for 14.2.2.1. It delivered major upgrades in perception, routing, and end-of-trip behavior for Full Self-Driving (Supervised).

The update introduced a higher-resolution vision encoder neural network that refines how the system reads the scene around the car. This encoder gives FSD better sensitivity to emergency vehicles, road debris, and human gestures, which is key for situations like a police officer directing traffic by hand. The improvement leans on Tesla’s Hardware 4 platform, which typically offers more compute and memory than Hardware 3, including larger RAM and storage capacity.

This upgrade links to new behavior on the road. FSD can now recognize flashing lights from emergency vehicles and either pull over or slow down, depending on context. Videos from owners show cars moving to the side of the road in response to approaching police or fire vehicles without an explicit driver command.

Arrival options and parking behaviors

A headline feature in v14.2.2 is “Arrival Options,” a set of choices that lets drivers tell FSD how they want the trip to end. Owners can select Parking Lot, Street, Driveway, Parking Garage, or Curbside, and the navigation pin will shift to an appropriate final point for that choice. For repeat destinations like home or work, these preferences can stick so the car uses the same arrival type each time.

Vision-based detours and integrated routing

Another key change is the deeper integration between the vision network and routing logic. FSD v14.2.2 can spot blocked roads, construction zones, or closures with the camera system and then choose alternate paths in real time.

Instead of running separate stacks for planning and vision, v14 pulls more of the navigation behavior into the end-to-end neural network.

Speed profiles and “Mad Max” return

FSD v14.2.2 also updated its Speed Profiles. Owners now have more granular options, from very conservative to very aggressive behavior. The lineup includes “Sloth,” “Chill,” “Standard,” and “Hurry,” and once again brings back “Mad Max” mode.

The v14 line is a major technical step for Tesla. The company has said the main driving network is about ten times larger in parameters than in v13, which allows more nuanced decisions at the cost of greater compute demand.

To keep this large model running on car hardware, Tesla uses a mixture-of-experts structure. In this design, different parts of the network specialize in particular driving tasks, such as city streets, highways, or parking lots. A gating component chooses a small subset of these expert networks to run at each moment, instead of powering the entire model at full capacity.

Hardware 4 operates around a 200-watt envelope, so smarter allocation of compute keep the system within thermal and electrical limits, even as the neural network grows.

Unified end-to-end policy

Earlier generations of FSD used different code paths for highway driving, surface streets, and parking. Those handoffs could produce uneven behavior when the car moved from one environment to another. By contrast, v14 applies a single policy network across all scenarios, including parking lots and garages.

Route planning is now more deeply tied to the same vision-based policy that manages steering and throttle. When the car encounters an unexpected closure, this integrated network updates both the path and the driving behavior, rather than passing control to a separate planner module.

FSD v14 training relies heavily on reinforcement learning in simulation as well as large-scale real-world data. This training approach pushes the system to optimize for good outcomes instead of simply copying average human inputs.

You may also like to read: