XPeng, in collaboration with Banma Intelligence, which is supported by Alibaba and SAIC Motor, has developed an automotive solution for an intelligent cockpit capable of running a large 30-billion parameter AI model directly within the vehicle.
The system was unveiled at the Alibaba Cloud Apsara Conference in 2026. It combines XPeng's proprietary Turing chip, which boasts an efficient computing power of 750 TOPS per chip, with Banma's multimodal onboard model, AutoOmni.
The design concept differs from the traditional separation of functions between the car and the cloud. The partners refer to this as an edge-oriented architecture with cloud support. Core functions requiring high frequency are executed locally: continuous multi-turn dialogue, multimodal perception and interaction, answering questions from a local knowledge base, and complex task planning. The cloud is only used for secondary, low-frequency functions such as news and weather.
This approach allows the cockpit to function with minimal latency in tunnels, underground garages, and other areas with poor or no signal, and it also addresses network coverage issues abroad or privacy requirements, according to the partners' statement.
This represents a significant leap in scale: according to a Xinhua report from the Banma forum at the conference, most existing cockpit onboard models have only 4 or 7 billion parameters. Michael, head of XPeng's intelligent computing center, noted that competition in the large model field is shifting from counting parameters in the cloud to inference on the device, and this solution demonstrates stable operation of a 30B class model on an automotive chip.
The Turing chip is designed to serve three of XPeng's product lines simultaneously: AI vehicles, AI robots, and flying cars. Banma provides the operating system and the model layer. The company claims that over 60% of automakers who have integrated onboard cockpit models into production use its solutions. At the same conference, Banma introduced AutoOmni 2.0-23B-A3B, a mixture-of-experts model with approximately 3 billion active parameters. Banma stated that this model handles routine cockpit tasks as well as cloud models ten times larger.
Together, the companies offer automakers a comprehensive solution that includes one chip, an AI operating system, and a large model, instead of separate components. Hao Fei, co-founder of Banma, reported that partners will start with smart cars and then expand joint hardware and software AI solutions to other embodied intelligent devices. They plan to continue adapting the model and improving tooling, focusing on mid-to-high-end vehicles. Neither company named a specific production model or launch date.
