Why is Chery an AI company?

Author | Wang Ruihao Editor | Li Yuchen

"Several leaders told me they feel Chery's cars are looking less and less like cars, and Chery is looking less and less like a car company. That's exactly what I want to hear—Chery is a high-tech innovation company 'wearing a car company's vest.'"

On October 20, the 2026 Chery Global Innovation Conference opened in Wuhu, Anhui. The person who said this is Yin Tongyue, chairman of Chery Automobile, who has spent over forty years in the auto industry. A man who has given most of his life to cars voluntarily taking off the "car company" signboard is not modesty—it's a judgment.

His basis is a metaphor: in 1886, Benz built the first car, and those left behind by the times were the horse keepers who had spent their lives driving horse carriages. They did nothing wrong; they just didn't see that the road had already changed. "The real risk is not that AI is too strong or too fast, but that we cannot continue to be the horse keepers of that era," Yin Tongyue said. "What AI will truly eliminate in the future are companies that reject AI and dare not use AI."

Leiphone.com believes that judging whether Chery is an AI company cannot depend on how many times it shouts AI at its launch events, but on three things: whether AI has restructured its production processes, whether it has settled into a reusable unified foundation, and whether it has grown new species beyond cars.

General-purpose large models on the market are like an encyclopedia—they can chat about anything, but once they enter the factory workshop they are no longer enough: insufficient precision, insufficient depth, insufficient reliability. Chery's choice is to develop its own vertical foundational large model, CheryMind, which has three major characteristics: a brain that understands the physical world, co-evolution with the business, and deployability across all domains.

What supports it is not crawled internet corpora, but the family assets Chery has accumulated over nearly 30 years: 700TB of enterprise knowledge, covering technical standards, process specifications, design guidelines, and safety specifications; 200PB of engineering data, from simulation results to test data, from CAD to the full chain of real vehicles; tens of millions of R&D documents; 25 million entries of coach-level driving data; 5 million entries of robot motion capture data.

The barrier in the physical AI era is not how much data there is, but where the data comes from. General-purpose large models crawl the internet; CheryMind feeds on the technical standards, process specifications, and simulation and test data accumulated over 30 years of carmaking. This is teaching material that others cannot buy and cannot crawl away.

Leiphone.com has learned that it is already running, not a PPT: 1,500+ end-to-end business processes, 400+ enterprise systems connected, 60,000+ daily uses by employees, and 200+ business scenarios covered. Gao Xinhua, executive vice president and CTO of Chery Automobile, gave a comparison on site: previously an engineer doing one round of simulation optimization had to run 6 rounds and spend hundreds of days; now, based on CheryMind, each person can produce 100 plans with results in half a day, a 42-fold efficiency improvement.

CheryMind walks on two legs. Enterprise-side vertical models make AI understand carmaking better: CAD vertical models, quality vertical models, manufacturing vertical models, and enterprise Agent vertical models, covering the entire chain of R&D—manufacturing—quality—supply—marketing—operations; product-side vertical models make cars understand people and the physical world better: smart cockpit, smart driving, vehicle-side intelligence, flight control vertical models, MoAgent, and agricultural machinery OS, extending all the way to eVTOL, robots, and agricultural machinery.

The key lies in the closed loop: new capabilities formed by vertical models in real scenarios continuously flow back to the foundation model. The two engines of business and product feed each other, and CheryMind grows stronger the more it is used. Self-evolution is the hardest criterion for judging whether one is an AI company—tools do not self-evolve; only a foundation does.

Beyond CheryMind, Chery released self-developed chips, named after Ziwei Star—the North Star, surrounded by stars, guiding the direction. The Ziwei chips have three major series, each handling one thing.

The C series handles thinking. The central computing chip C100 has 2,400TOPS of vehicle NPU computing power, 600GB/s of memory bandwidth, supports 100B large-model on-device deployment, and completes the E2E algorithm closed loop from perception to decision in 80ms—less than the time of a human blink. Benchmarking against industry flagships, it has 2.2x the bandwidth, 1.2x the computing power, and 1.5x the image processing.

The S series handles energy. The automotive-grade silicon carbide S1500 has a withstand voltage of 1,500V, the highest in the industry; an output current of 1,450A, efficiency ≥99.6%, a junction temperature of 200°C, and mass production on 8-inch wafers cuts costs by more than 30%.

The M series handles control. Automotive-grade control chips: MC200 domain control, MS100 pre-driver, MX100 PMIC, and MC100 electric drive, to be implemented in sequence starting from 2027.

Behind these three things is Chery's account over the past five years: R&D investment exceeding 110 billion yuan, 313 Yaoguang laboratories, and smart-driving penetration up 267% in two years. What AI rewrites is the cost curve of carmaking itself, not marketing rhetoric.

More noteworthy is a layer of synergy: large models define what chips should look like, and 100B on-device deployment forces 2,400TOPS computing power and 600GB/s bandwidth; chips in turn determine how deep the models can land. This kind of "model—chip" co-definition is a characteristic of tech companies, not integrators.

What this conference truly released is not a single technology. Yin Tongyue's summary is: mobility, energy, food, and life are not four isolated directions; they share models, chips, data, and engineering capabilities, and continuously generate new demands in real scenarios. The structure is "one foundation × N scenarios."

The foundation is first used in the hardest scenario: smart driving going global. In Southeast Asia, two-wheelers and cars share lanes, dark intersections have wrong-way driving, forced lane cutting... Solutions that work in the Chinese market may not hold overseas.

Chery's judgment is: in global competition, what should be contested is not computing power, data, or the number of lidars, but the model foundation. Without a unified foundation, every entry into a new market and every launch of a new car requires redoing a whole set of solutions.

Chery's All-Dimensional World Engine (CHERY AWE) uses three models to handle this.

The Structuring World model is responsible for "understanding the world": it has built 700,000+ scenes, including 300,000+ high-risk and 20,000+ extreme ones, uniformly encoding five types of data—multi-view images, LiDAR point clouds, scene structure, descriptive text, and driving trajectories—into latent space to support closed-loop simulation of long-tail scenarios; the Perceiving Language model is responsible for "understanding semantics": multimodal reasoning plus short-term memory and a language topology graph, reaching a destination using only road signs like a veteran driver, and outputting a traceable reasoning chain; the Action Policy model is responsible for "producing behavior": a MoT multi-expert architecture, shifting from "passively responding to danger" to "actively predicting risk".

The Falcon assisted driving system has already run with this foundation for a year: meeting regulatory access in 130+ countries and regions, earning C-ICAP five-star+ and IVISTA G+ certifications, and winning six consecutive championships in the China Intelligent Driving Competition.

Falcon intelligent driving comes in three versions: Falcon 500 offers urban capability from the entry level, with equal intelligence for gasoline and electric; Falcon 700 covers full-scenario parking-spot-to-parking-spot and cross-floor valet parking; Falcon 900 covers highway and urban L3, with full redundancy in perception, control, and power.

On the cockpit side, the Lingxi 8D intelligent sensing cockpit uses the AIOS full-domain architecture to connect the five domains of chassis, cockpit, body, intelligent driving, and powertrain, with vehicle-level signal coordination and millisecond-level cross-domain response. Outside the cabin it recognizes pedestrian intent; inside it senses emotion and fatigue, upgrading from "people finding functions" to "functions finding people"; hearing, vision, smell, motion, touch, emotion, health, and social interaction are orchestrated across eight dimensions, with one-sentence Skill free orchestration and personalized experiences for every individual. Details such as Boya Sound with 33 speakers, a 43.7-inch PHUD, and S2S end-to-end voice all push toward "the car understanding people better."

The industry's first native front-installed Robotaxi born for the globe: five natives—architecture, front-installation, experience, operations, and globalization—and six layers of safety—passengers, behavior, perception, algorithms, execution, and fleet—with a goal of 100,000 vehicles in global operation by 2030.

Robots share the same origin as vehicles: 30 years of automotive supply chain, hardware architecture, and quality systems are directly reused, with 31 innovation application laboratories laid out across six major fields. Under the Multi-agent intelligent architecture, MoAgent 2.0 developed based on CheryMind uses a slow brain for understanding and planning and a fast brain for real-time response, with device-edge-cloud collaboration forming a data loop of "execution—feedback—optimization—re-execution."

The product matrix is already in place: quadruped/wheeled robot dogs, smart police, medical guides, Mobao family companionship, industrial and humanoid robots; Moja robots have now shipped 2,000+ units overseas, covering 60+ countries.

In addition, in terms of energy freedom and food freedom, Xunlong instant charging provides 500 kilometers of range in 5 minutes, and the solid-state battery pilot line will be put into use in the first half of 2027; smart agricultural machinery reduces energy consumption by 74%, raising per-mu yield on the same land from 1,600 jin to 2,000 jin, with the goal of serving 1 billion mu of farmland globally by 2035.

Putting these scenarios together reveals Chery's most unique lever in doing AI: reuse rate is the gross margin of the AI era. The same set of models, chips, and supply chains migrates across automobiles, Robotaxis, robots, and agricultural machinery, with diminishing marginal costs. This is an advantage only a manufacturing-origin company has in doing AI.

Chery also clearly knows it cannot run all scenarios alone. It is moving from product manufacturer to ecosystem organizer: opening technology platforms, scenarios, and industrialization capabilities to suppliers, universities, and entrepreneurs, and co-building the 2050 Future Academy with USTC to pursue long-cycle, fault-tolerant frontier exploration.

In Yin Tongyue's words: "Technology can only truly create freedom when it leaves the laboratory and enters roads, homes, farmland, and energy systems, continuously solving real problems."

Chery at thirty no longer wants to be just an automaker. That sounds like a slogan, but when an automaker's robots begin to truly solve social pain points and its models begin to determine what its own chips look like, it is no longer just a slogan.

Chery's New Battle: Going Global, Intelligent Driving, and Head-On Confrontation with BYD and Others

Bigger than a Mercedes-Benz S, can the Han EV support BYD's ambition to move upmarket?