XPeng and Banma’s 30B-Class On-Device AI Cockpit: What the Announcement Does—and Doesn’t—Show

XPeng and Banma announced a joint AI cockpit design that supports running 30-billion-parameter-class models on the vehicle, using XPeng’s Turing chip, which the announcement describes as providing up to 750 TOPS of computing power per chip. The approach puts local processing first and uses cloud services as support. The announcement does not name a production vehicle adopting the joint system or give a launch date. 61821

What the two companies are contributing

XPeng brings its Turing in-vehicle compute platform; Banma brings its AutoOmni multimodal cockpit model. The partners describe the system as supporting local 30B-class model operation and functions including multi-turn conversation, multimodal interaction, offline question answering and task planning. 1821

The announcement is part of a broader effort to offer vehicle technology beyond XPeng’s own cars: Reuters reported that XPeng plans to offer its chips, cockpit systems and other technology to other automakers. Banma, meanwhile, is described as a smart-car solution provider backed by Alibaba Group and SAIC Motor. 15

What “on-device first, cloud-assisted” means

In this design, supported cockpit tasks can be processed in the vehicle rather than routinely sent to a server. The reported use cases include offline Q&A, while cloud services remain available for some lower-frequency needs. That could make supported functions less dependent on an internet connection and avoid network round trips for local tasks. 1821

Those are potential architectural benefits, not measured results. The available announcement does not provide latency tests, describe specific cabin-data handling practices, or establish how the system performs against particular 4B- or 7B-parameter cockpit models. A larger parameter count alone does not prove that a model will be faster, more capable or more private in a vehicle; those comparisons require testing and implementation details that have not been provided here.

The separate AutoOmni 2.0-23B-A3B model

Banma also announced AutoOmni 2.0-23B-A3B, a separate on-device model built using a mixture-of-experts architecture and described as having 23 billion parameters. Banma says it performs comparably to 230B-parameter cloud models on routine cockpit tasks and reaches 90% of similarly sized cloud models’ performance on complex tasks. These are vendor-reported comparisons, not independent head-to-head results. The available sources do not establish that the 23B model is the same configuration as the 30B-class model in the XPeng-Banma joint solution. 4

What is still unconfirmed

The announcement describes a joint solution, but the available reports do not identify a production vehicle using it or provide a deployment timeline. Nor do they confirm that the joint cockpit system will be deployed in a robot or another non-car device. XPeng has broader plans to offer technology to other automakers and expand licensing into robotics and other physical-AI applications, but those plans are distinct from a confirmed deployment of this cockpit solution. 16

The practical test will be whether automakers adopt the system and how it performs in real vehicles. Until those details emerge, the 30B-class figure and on-device-first design describe the announced capability and approach—not a verified advantage over smaller cockpit models.