Banma AutoOmni 2.0 Brings Omni-Modal AI to Smart Cockpits

Banma Intelligence of Alibaba Group Holding Limited (Alibaba Group) announced the release of Banma AutoOmni 2.0, an omni-modal on-device AI model identified as AutoOmni 2.0-23B-A3B. The company also demonstrated an in-vehicle implementation of its AutoClaw 2.0 smart cockpit collaborative service solution. Together, the two offerings address how artificial intelligence can be delivered inside the vehicle cabin, with the model running on-device and the solution combining on-device and cloud capabilities to support smart cockpit functions.

Mixture-of-Experts Model Targets Cloud-Level Cockpit Performance

AutoOmni 2.0 uses a mixture-of-experts (MoE) architecture. According to the company, its performance on routine smart cockpit tasks can match that of a cloud-based model with ten times as many parameters. For complex tasks, it achieves 80%–90% of that model's performance. These figures are company-reported, and they position the on-device model as a close alternative to a much larger cloud-based system for everyday in-cabin requests, while the gap widens somewhat on more demanding tasks.

AutoClaw 2.0 Combines On-Device and Cloud Agents

The AutoClaw 2.0 in-vehicle solution uses an integrated on-device and cloud agent architecture. It can handle ambiguous user intent, navigation involving multiple tasks, vehicle control, multimedia and cross-domain tasks. This range of capabilities covers several of the core functions of a smart cockpit within one collaborative service solution, which Banma Intelligence showed in an in-vehicle implementation as part of the announcement.

Chip Partnerships Drive Speed and Memory Gains

Banma Intelligence said it has partnered with ten domestic and international chip companies on AutoOmni. Hardware and software co-optimization enables stable concurrent processing of more than eight tasks, delivers a five- to sixfold increase in model inference speed, maintains quantization fidelity above 99%, and reduces runtime memory usage by more than 50%. Because on-device models must run within the limits of in-vehicle hardware, these results indicate the focus of the optimization work carried out with the chip partners.

Industry Impact & Outlook

The release reflects a broader push to run capable AI models directly in the vehicle rather than depending solely on the cloud. If the reported results hold in production, smart cockpit developers could gain a path to strong assistant performance within on-device hardware limits, which matters for automakers and suppliers weighing cost, responsiveness, and system complexity. The partnerships with ten chip companies suggest Banma Intelligence is working to make the model adaptable across multiple hardware platforms, an important factor for adoption. Next steps to watch include independent validation of the performance claims, announced vehicle programs, and how competing cockpit AI providers respond.

Frequently Asked Questions

What is AutoOmni 2.0-23B-A3B?
AutoOmni 2.0-23B-A3B is an omni-modal on-device AI model released by Banma Intelligence of Alibaba Group Holding Limited, built on a mixture-of-experts architecture for smart cockpit applications. According to the company, its performance on routine smart cockpit tasks can match that of a cloud-based model with ten times as many parameters. For complex tasks, it achieves 80%–90% of that model's performance. The two products were presented together as part of the same announcement.

What can AutoClaw 2.0 do inside a vehicle?
AutoClaw 2.0 is an in-vehicle smart cockpit collaborative service solution that uses an integrated on-device and cloud agent architecture to process user requests. It can handle ambiguous user intent, navigation involving multiple tasks, vehicle control, multimedia and cross-domain tasks. Because the architecture integrates on-device and cloud agents, requests that span several functions can be coordinated within a single solution. Banma Intelligence demonstrated the in-vehicle implementation during its announcement of the AutoOmni 2.0 model.

What performance gains come from hardware and software co-optimization?
Banma Intelligence said hardware and software co-optimization enables stable concurrent processing of more than eight tasks, following its partnerships with ten domestic and international chip companies. The company also reported a five- to sixfold increase in model inference speed. Quantization fidelity remains above 99%, and runtime memory usage falls by more than 50%. These figures are company-reported and reflect the optimization work done with its chip partners. Together, they target stable performance on in-vehicle hardware.