A driver tells the car, in plain language, to pull over by the white car ahead — and the car does it, without a programmer having written a rule for that exact spot. That is the promise Li Auto (理想汽车) is shipping, not demoing. Its driver model is now one of the most-used large-model features in any Chinese vehicle.
Key takeaways
- Li Auto's VLA driver model (VLA司机大模型) reached 80% monthly usage among users and logged 12.254 million voice commands by the end of 2025.
- The company released MindVLA-o1, its next-generation autonomous-driving foundation model, at NVIDIA GTC 2026 on 17 March 2026.
- VLA stands for Vision-Language-Action (视觉—语言—行动): one model that perceives, reasons and acts, rather than separate perception, prediction and planning modules.
- Li Auto pushed the VLA driver model to all AD Max users in September 2025 after an August rollout with the Li i8.
- The strategy is explicit: Li Auto calls itself an "All in AI" physical-AI company with the car as its largest robot.
From rules to one model
Legacy assisted driving is a stack of separate modules — perceive, predict, plan — each hand-tuned. Li Auto's bet is to collapse that into a single Vision-Language-Action (视觉—语言—行动) model. The "language" part is what lets a passenger issue a fuzzy instruction and the car interpret it; the "action" part is the driving output. The company says this is less a feature than a shift toward a general agent for the physical world.
The rollout was staged. Li Auto released its self-developed cognitive model Mind GPT in June 2023, pushed an end-to-end plus VLM (端到端+VLM) architecture in October 2024, and in 2025 unified spatial understanding, language and action into the VLA driver model, shipping it with the Li i8 in August and to all AD Max users in September. By year-end the model was being used at scale: 80% monthly user rate and 12.254 million accumulated commands, with 1.303 million commands during the Spring Festival period alone across 250 million kilometres of assisted driving.
MindVLA-o1 at GTC
At NVIDIA GTC 2026 on 17 March, foundation-model lead Zhan Kun (詹锟) unveiled MindVLA-o1, described as a native multimodal Mixture-of-Experts Transformer. Li Auto lists five technical pillars: 3D spatial understanding, multimodal "thinking," unified action generation, closed-loop reinforcement learning, and hardware–software co-design. The pitch is that the same VLA model can drive a car and, in principle, extend to a robot — "autonomous driving is only the start of physical AI."
According to Li An, chief scientist at BrainNet (脑机网), unifying vision, language and action in one model is where Chinese automakers are trying to leapfrog the modular stack, not merely catch up — the advantage is iteration speed, not any single benchmark.
The research footprint backs the claim modestly: Li Auto says MindVLA-o1-related work appeared at venues including CVPR and ICLR in 2026. Those are the company's own citations and should be read as evidence of serious investment rather than independent validation of performance.
Why a carmaker becomes an AI company
Li Auto's public framing is no longer "car manufacturer." It describes itself as an "All in AI" physical-AI company with the automobile as its largest robot. The logic is that a car is the most complex physical product most consumers own, so mastering driving with a general model is a stepping stone to general embodied intelligence. That reframing also changes what the company hires for and what it tells investors it is worth.
The business upside is concrete. A model that understands natural-language commands turns the cabin into a controllable assistant and opens agent-style features — booking, parking payment, routine automation — that stick users to the brand. High usage (80% monthly) is the kind of engagement metric that justifies further model spend.
The caveats that matter
A VLA model is only as good as its failure behaviour. Language grounding can fail on rare scenes, and a single unified model concentrates risk: a reasoning error can propagate straight to steering. Li Auto's own history notes earlier end-to-end systems needed work on sudden scenarios. The 12.25-million-command figure counts interactions, not disengagements or safety outcomes, so it measures adoption, not capability.
Benchmark claims here are largely the company's. We have not independently tested MindVLA-o1's driving quality, edge-case handling, or real-world disengagement rate. The "physical AI" vision is a strategy statement, not a shipped robot product.
Honest limitations
This article draws on Li Auto's own community and WeChat posts, state-media reporting (CCTV, CRI, Beijing Daily) and the World Internet Conference profile. The 80% usage and 12.254-million-command figures are company-reported for end-2025 and not independently audited; the 250-million-km Spring Festival distance is a company statistic. MindVLA-o1's architecture claims are Li Auto's presentation at GTC 2026. We did not verify real-world disengagement rates, safety cases, or third-party benchmark rankings, and the "first/leading" framing is the company's.
What readers can do now
- Treat the 12.25-million-command number as adoption, not capability; watch for disengagement or safety data before judging the model.
- For competing carmakers, the lesson is organizational: a unified model needs a foundation-model team, not just a perception team.
- For users, try vague, location-based voice commands in a safe setting to see where the VLA model helps and where it still defers to driver control.