HW4 vehicles run distilled versions of Tesla’s FSD models

HW4 vehicles run distilled versions of Tesla’s FSD models built for next‑gen hardware

As Tesla works toward unsupervised self-driving across multiple hardware generations, an important detail has surfaced: customer cars with Hardware 4 (HW4/AI4) are not running Tesla’s base full self-driving model.

Instead, the FSD builds on HW4 are scaled-down versions of networks designed for the company’s next-generation platform. Development is centered on the Cybercab first, and those models are then distilled to operate on the slightly less powerful HW4 systems used in current production vehicles.

The Cybercab includes more powerful FSD hardware than today’s consumer cars. That additional compute gives Tesla a larger sandbox for training and exercising heavier, more complex AI networks that can later be streamlined for the wider fleet.

What is Distillation?

Running a massive neural network inside a moving vehicle demands significant processing power and memory. To bridge the gap between data center training and in-car execution, Tesla relies on a technique known as distillation.

Engineers train large, complex “teacher” models in data centers. After these models learn to handle difficult driving tasks, they are distilled into smaller “student” models that are optimized to run locally on consumer vehicles, offering similar capabilities without overloading the on-board hardware.

The degree of distillation depends on the target hardware, particularly memory capacity and bandwidth. While the Cybercab represents the ideal hardware setup for now, customer vehicles receive optimized, lighter variants.

Comparing the Hardware Gap

This mirrors how Tesla created FSD v14 Lite for older Hardware 3 (HW3/AI3) vehicles. The v14 Lite build is a distilled version of the main v14 branch that started reaching early access testers last month.

The FSD v14 builds running in customer vehicles are likewise distilled from what operates natively on the Cybercab. HW4 likely does not require nearly as much optimization as HW3, which according to Tesla only has about 15% of the memory bandwidth of its successor.

Although full specifications of the Cybercab’s next-generation processor remain undisclosed, it is known to feature more RAM than HW4, which is critical for AI. It has been speculated that the Cybercab may be using the new AI4+ chips Tesla announced during its Q1 earnings call earlier this year. Those chips are expected to come with 64GB of memory or potentially even more. In addition to a more capable computer with greater memory, the Cybercab also includes dual GPS for more precise location tracking during driverless trips.

Unsupervised Hopes Stay Alive for HW4

This does not mean HW4 will be unable to achieve unsupervised FSD. Tesla has previously confirmed that HW3 won't be able to achieve unsupervised autonomy (owners have been promised hardware upgrades that will make their vehicles capable of achieving it), but HW4 remains in contention. The distillation approach is encouraging for HW4 owners—running a version of the software being developed to power fully driverless, commercial operations on the Cybercab places HW4 in a strong position to reach true autonomy.

By targeting the most powerful hardware first and then distilling downward, Tesla can fully utilize the Cybercab’s capabilities while keeping customer cars capable. In practice, the FSD builds on consumer vehicles are optimized, lighter versions of the company’s more advanced robotaxi software.