--- title: "Li Auto's MindGPT makes the car cabin an agent, not a chatbot" date: 2026-11-01 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-li-auto-mindgpt language: en --- # Li Auto's MindGPT makes the car cabin an agent, not a chatbot > Li Auto's in-car assistant 理想同学 runs on its self-developed MindGPT base model, now an end-to-end agent, while a separate Vision-Language-Action model helps steer the car. You ask your car where to stop for coffee that fits a toddler's nap schedule, and it books the stop, adjusts the cabin, and remembers the preference next time. That exchange is now routine for over a million Li Auto owners — and it runs on a model the carmaker built itself rather than licensed from a general-purpose lab. Li Auto's cabin assistant, 理想同学, is powered by MindGPT, a base model the company has iterated since 2023. In 2025 the assistant was rebuilt as an end-to-end agent that can plan, call tools, and search while it thinks. A separate model, built on Vision-Language-Action (视觉—语言—动作) ideas, began rolling out to driver-assistance hardware the same year. Together they show a path where the car, not the phone, becomes a primary place people meet a foundation model. ## Key takeaways - Li Auto's in-car assistant 理想同学 is powered by the company's self-developed MindGPT base model, upgraded to an end-to-end agent in 2025. - A separate Vision-Language-Action (视觉—语言—动作) driver model began full rollout to driver-assistance hardware in September 2025. - Li Auto's 2025 ESG report shows assisted driving crossed 3.07 billion km and the cabin assistant logged 2.38 billion wake-ups. - MindGPT cleared China's generative-AI service filing among the first carmaker models to do so. ## From voice assistant to agent Early versions of 理想同学 were a voice assistant: answer a question, set a navigation, play a song. The 2025 upgrade changed the shape of the interaction. MindGPT 3.1 is described by the company as an end-to-end agent model with "think-while-search" behavior — it can call tools, query real-time information such as tickets, stock prices, or exchange rates, and handle multi-step tasks without the user chaining prompts by hand. The company says the model was trained on more than 3 trillion tokens across driving, entertainment, and travel scenarios, using a TaskFormer architecture that combines Mixture-of-Experts routing with sparse attention. The practical payoff is less about a bigger brain and more about reliability: the assistant can ask a clarifying question when a request is vague instead of guessing. ## The model that also drives The more striking deployment is outside the cabin. Li Auto's VLA (Vision-Language-Action) driver model — a "司机大模型" — fuses visual perception, language understanding, and driving decisions into one system. The company reports it reached full rollout to its AD Max hardware in September 2025, improving route choice, speed strategy, lane changes, and parking comfort through reinforcement learning. During the 2025 National Day holiday period alone, the company logged 1.817 million hours of assisted driving and 1.225 million uses of VLA commands. The promise is a driver model that, like the cabin assistant, keeps improving the more it is used. ## What the numbers say Li Auto's 2025 ESG report gives a rare, official window into scale. Assisted driving totaled 3.07 billion km and 37.29 million hours; smart parking was triggered 130 million times; the cabin assistant recorded 2.38 billion wake-ups with a 96% task-completion rate and 96.2% satisfaction, serving roughly 1.5 million owners by year-end. Those are deployment numbers, not benchmark scores, and they matter more for judging a carmaker's model than any leaderboard. They show the model is not a demo — it is running continuously across a large fleet, generating the feedback loop that improves both the cabin agent and the driver model. ## Why carmakers build their own models A car is a privacy-sensitive, safety-critical, always-on environment. Relying on a third-party cloud model for every interaction creates latency, cost, and data-residency problems. Building MindGPT in-house lets Li Auto keep personal preferences, cabin context, and driving perception inside its own stack, and tailor the model to the odd, specific vocabulary of life in a vehicle. It also creates a moat. The assistant that learns a family's seating, climate, and entertainment habits becomes harder to leave than any single feature, and the driver model's improvement compounds with fleet data the company alone collects. ## Honest limitations The deployment figures come from Li Auto's official 2025 ESG report, which we treated as authoritative; we did not independently audit them. Specific model claims — 3 trillion training tokens, the TaskFormer architecture, "think-while-search" speed gains — come from the company's own materials and were not reproduced. The VLA driver model supports assistance, not autonomous driving, and its real-world safety depends on hardware, sensors, and oversight we did not test. We did not benchmark MindGPT against general-purpose models on standard suites. ## What readers can do now - If you own or shop for a Li Auto, try 理想同学 with a multi-step request (book, adjust, remember) and judge whether the agent behavior actually saves you taps. - If you build in-vehicle AI, study the cabin-plus-driving split: a private on-vehicle model for habits, a perception-action model for the road. - If you track the field, watch whether other carmakers follow Li Auto in clearing China's generative-AI filing and shipping agent-class assistants at fleet scale. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-li-auto-mindgpt Free to quote with attribution and a link to the original.