Li Auto VLA Driver Model

What it is

VLA Driver is Li Auto's driver-assistance model for its passenger cars. The name stands for vision, language and action, as Lang Xianpeng, Li Auto's senior vice president of autonomous driving research and development, described it.^[3] In plain terms, the model is meant to take in what the car's sensors see, work with instructions given in ordinary language, and turn both into decisions about how the vehicle should move. Building the driving system around one model rather than a stack of separately written rules is what the name is drawing attention to.

How it reached customers

Li Auto began delivering the VLA Driver model together with the Li i8 on 20 August 2025.^[1] The company said every existing owner of a car with AD Max hardware would be upgraded to the VLA Driver model, so the new software was not reserved for people buying a new car.^[1] It also said the model would run on a wide range of hardware: both the 2022 Li L9, which sits on the Orin X platform, and newer cars built on the Thor platform can run VLA.^[1]

The broad rollout came through OTA 8.0. Li Auto said the September 2025 update would roll the VLA Driver model out to all AD Max users.^[4] That split the job in two: the model shipped first with a new vehicle, and the update carried it back to the cars already on the road.

What the model is meant to do

Li Auto describes two behaviours that separate VLA from a fixed-rule driver-assistance system.

The first is reasoning about what the car cannot see. Li Auto says VLA can reason about blind spots at T-junctions and brake before a potential collision, instead of applying a fixed rule.^[2] A fixed rule can only fire when its trigger condition is met; a model that reasons about the situation is meant to notice that something could be coming from a direction the sensors do not cover, and slow down on that basis. Junctions where a road meets another at an angle are a standard hard case for this, because a car approaching along the crossing road is hidden until late.

The second is conversation. Li Auto says VLA supports continuous, multi-turn task conversations, comparable to talking with a human driver.^[2] Rather than issuing one command per action, a driver can keep talking with the system, adding and adjusting instructions across several turns.

Where it sits in Li Auto's lineup

VLA Driver is the AD Max side of the story. Li Auto also said that AD Pro would receive a city NOA function based on the Horizon Journey 6M chip by the end of 2025,^[2] so the two hardware tiers move on separate schedules with different chips and different software.

OTA 8.0 carried more than the driving model. The same update introduced the Li Xiang Tong Xue agent, which Li Auto describes as moving from a voice assistant to an agent that can use tools.^[4] The two changes point in the same direction: an assistant that acts on a request rather than only answering it, and a driving system that reasons about a situation rather than only matching it against a rule.

Why it matters

The VLA Driver release is a case of a carmaker shipping a driving model to a fleet that already exists, across two generations of compute hardware, rather than tying it to a single new product.^[1][4] For owners, the practical question is what the model does at the moments rules handle badly — an occluded junction, a request that changes halfway through. Li Auto's own description answers that question in terms of reasoning and conversation.^[2]