Banma Intelligence unveils AutoOmni 2.0 on-device AI model

Banma Intelligence Group has introduced AutoOmni 2.0-23B-A3B, an omni-modal on-device large language model (LLM), at the Apsara Conference. Utilizing a mixture-of-experts (MoE) architecture, the model is designed to perform routine intelligent cockpit tasks at levels comparable to cloud models ten times its size. For more complex operations, it reaches approximately 90% of the performance of similarly sized cloud models.

According to Chief Technology Officer Dr. Si Luo, the model is compatible with most mainstream high-performance automotive SoC platforms. It supports the concurrent execution of over eight multitasking streams and accelerates large-model inference by five to six times. The technology aims to create a “native intelligence era” for vehicles by combining on-device processing with a cloud-integrated agent architecture to handle ambiguous voice commands and cross-domain tasks with low latency.

Demonstrations at the conference included the FREELANDER 8, a vehicle co-developed by Chery and Jaguar Land Rover, which is the first production vehicle to feature AutoOmni on Qualcomm’s Snapdragon Cockpit Elite platform. Other showcased models included SAIC’s MG Cyberster and BYD’s DENZA Z9GT.