The longest commute in China this year was not a person — it was a neural network, looping through billions of kilometers of real roads.
A carmaker that once competed on range and seating now ships software that decides when to change lanes and where to park.
The next frontier is not another model but the silicon underneath it, and Li Auto just built its own.
Key takeaways
- Li Auto (理想汽车) said its VLA Driver large model (VLA司机大模型) reached full rollout to AD Max vehicles in September 2025, after a first push in August.
- In 2025 the fleet added 3.07 billion km (30.7亿公里) of assisted-driving distance, with a 99.5% active-user rate; active safety avoided more than 12 million potential accidents.
- The company reported 2025 R&D spend of RMB 113亿 (≈ US$1.56B), about 10.1% of revenue, with AI-related work roughly half of that.
- Its self-developed 5nm driving chip, Macha 100 (马赫100), taped out in 2025 and is slated for deployment in 2026; its Halo OS (星环OS) has been open-sourced since April 2025.
The model that actually drives
Li Auto's annual results, filed to the Hong Kong exchange on 12 March 2026, describe 2025 as the year its assisted-driving (辅助驾驶) stack crossed from rule-based engineering into a learning-based system. The centerpiece is the VLA Driver large model (VLA司机大模型) — a vision-language-action architecture that fuses spatial, language and behavioral intelligence, with memory, reasoning and planning. The company says it began a full over-the-air rollout to AD Max models in September 2025, after an initial push in August.
Crucially, Li Auto claims a first: a mass-produced "world model + reinforcement-learning closed loop." In plain terms, it generates simulated driving scenarios in the cloud to train the on-car VLA model, which then improves from safety, efficiency and comfort feedback. That is a different industrial posture from simply shipping a model — it is a training pipeline attached to a fleet.
The distance that proves it
Numbers from the results and the company's 2025 ESG report:
- New assisted-driving mileage in 2025: 3.07 billion km (30.7亿公里).
- Total assisted-driving time: 37.29 million hours.
- Assisted-driving user active rate: 99.5%.
- Potential accidents avoided by active safety through end-2025: more than 12 million.
- During the 2025 National Day holiday alone: 1.817 million hours of assisted driving, 1.225 million VLA commands used, 3.26 million intelligent parking maneuvers.
These are fleet-scale figures, not benchmark scores, and they are the company's own. But they illustrate why a carmaker treats its driving model as core IP: the data loop is the moat, and the moat is measured in kilometers.
MindGPT rides in the back seat
While VLA handles the road, Li Auto's in-cabin assistant runs on its own foundation model, MindGPT (认知大模型). The company says the cabin base model was upgraded to MindGPT 3.1 in 2025, with faster response and new tool use — the "Li Xiang Tong Xue" (理想同学) agent can now call external services such as a KFC ordering agent, use the car's exterior cameras and a payment app to scan parking QR codes, and remember individual family members via Face ID. MindGPT was, per the company, the first carmaker-built model to clear China's generative-AI service filing.
Building the silicon
The most strategic line in the report may be the chip. Li Auto's self-developed 5nm autonomous-driving chip, Macha 100 (马赫100), completed tape-out in 2025 and has been road-tested on prototype vehicles, with deployment planned for 2026. A 5nm part designed in-house for the driving stack signals an intent to control both the model and the hardware that runs it — the same vertical-integration logic Tesla and Huawei have pursued, now attempted by a Chinese EV maker.
Alongside the chip, Li Auto open-sourced its vehicle operating system, Halo OS (星环OS), starting April 2025. By September 2025, 16 partners across the smart-car supply chain — including Great Wall Motor (长城汽车) and Infineon (英飞凌) — had signed the OS community charter, moving it into a jointly built phase.
According to Li An, chief scientist at BrainNet (脑机网), when a carmaker starts taping out its own driving silicon, the competitive question shifts from "whose model is smarter" to "who controls the cost per inference token at the edge. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网)."
What the spending signals
Li Auto's RMB 113亿 (≈ US$1.56B) R&D bill for 2025, about 10.1% of revenue with AI roughly half, is the financial backing for all of the above. The company's second-quarter 2025 revenue was RMB 30.2B (≈ US$4.2B), with net income of RMB 1.1B. The Li i8, priced at RMB 339,800 (≈ US$46,900), ships with the VLA Driver model and the Li Xiang Tong Xue agent as standard.
The picture is of a firm treating AI not as a feature add-on but as the product architecture: a driving model, a cabin model, a self-made chip and an open OS, all feeding one another.
Honest limitations
All quantitative claims are taken from Li Auto's 2025 annual results announcement (filed to HKEX on 12 March 2026) and its 2025 ESG report, both company disclosures. Assisted-driving mileage, active-safety counts and user-activity rates are the company's own metrics and are not independently audited. "Assisted driving" (辅助驾驶) denotes driver-support systems, not autonomous driving; the driver remains responsible. We have not benchmarked VLA against competitors' driving models, nor verified Macha 100's performance or 2026 production timing beyond the company's stated plan. Currency conversions use an approximate RMB/USD rate of 7.1–7.25 and are rounded.
What readers can do now
- Follow Li Auto's (02015.HK / LI) quarterly deliveries and any update on Macha 100 deployment as the signal that in-house silicon is real, not merely announced.
- Distinguish "assisted driving" from autonomy when reading the company's safety statistics — the driver is still in the loop.
- Compare Li Auto's vertically integrated model-chip-OS stack with peers such as Huawei, Xpeng and Tesla to judge whether in-house silicon becomes an industry norm.