Banma Intelligence launches AutoOmni 2.0 edge AI model, claims task performance comparable to cloud models with 10x parameters

Hangzhou, September 23 — During the Yunqi Conference, Banma Intelligence released its next-generation full-modal on-device large model, AutoOmni 2.0-23B-A3B, which delivers task processing capabilities comparable to cloud models with 10 times the parameter scale. The company also showcased the new AutoClaw 2.0 in-cabin collaborative service solution, designed to make AI better understand "my world."

Core Value of On-Device Intelligence: A New Interaction Paradigm

"The core value of on-device intelligence is to open a new paradigm of interaction," said Hao Fei, Co-CEO of Banma Intelligence. He emphasized that full-vehicle, always-on wake-free interaction—enabled by on-device models, along with contextual awareness and memory—are key elements of this new paradigm, making proactive, personalized services possible.

Autonomous Driving Understands the "Physical World"; In-Cabin AI Understands "My World"

At the Banma Intelligence booth during the Yunqi Conference, vehicles such as the Denza Z9GT, IM LS6, Shenxing 8, and Hongqi Tiangong 06 each featured different autonomous driving systems—Shenzhou Zhiyan, Momenta, Huawei Qiankun, and Zhuoyu, respectively. However, all chose Banma's Yuanshen AI for their in-cabin intelligence. According to the company, Yuanshen AI now covers 68% of services across mainstream automakers.

"Putting AI in the car does not mean winning over users; deep usage is the real battleground," said Cai Ming, Chief Product Officer at Banma Intelligence. He explained that for AI to perform complex planning and effectively orchestrate various capabilities, it must first understand the user. This involves two layers:

• The "Surface World": Understanding physical space, rules, and trajectories.
• The "Inner World": Understanding personal intentions, preferences, and life scenarios.

Through this dual understanding, AI can continuously evolve in a cycle of perception, comprehension, execution, and feedback.

According to Cai Ming, Yuanshen AI is evolving from an in-cabin AI toward a Personal AI, achieving "three connections":

1. Connecting cloud and vehicle in terms of computing power and services.
2. Connecting multiple user terminals, enabling cross-temporal and cross-location data integration.
3. Connecting users with services, enabling cross-app and cross-device service orchestration.

Xiao Ruizhe, founder and CEO of AI data company YoooClaw, stated that in partnership with Banma Intelligence, and under the premise of data compliance, the goal is to bridge the "context" between the user's one hour inside the car and 23 hours outside, fully unlocking the value of data and intelligence.

At the Banma booth, the new AutoClaw 2.0 in-vehicle solution, along with a new ecosystem agent solution co-developed with YoooClaw, demonstrated the initial implementation of these "three connections."

Smart Vehicles: The Largest "Pilot Plant" for Embodied Intelligence

During the forum, Banma Intelligence officially launched the AutoOmni 2.0-23B-A3B, a full-modal on-device large model based on a Mixture of Experts (MoE) architecture. In smart cockpit scenarios:

• For routine tasks, its performance is comparable to cloud models with 10 times the parameter count.
• For complex tasks, it achieves 80%–90% of the capability of such cloud models.

At the Yunqi Conference, Banma demonstrated a real-vehicle solution using this model and an end-cloud integrated agent architecture. The system handled fuzzy intent, multi-task complex navigation, vehicle control, multimedia, and cross-domain tasks with low latency, privacy protection, and stable, smooth execution.

AutoOmni was the industry's first full-modal on-device large model launched by Banma Intelligence. Within one year, 63% of automakers deploying on-device models have chosen Banma.

"Embed the on-device model into the chip, and embed safety into the core," said Si Luo, Chief Technology Officer of Banma Intelligence. He noted that AutoOmni has partnered with 10 chip companies globally, achieving:

• 8+ concurrent multi-tasks with stable performance through software-hardware collaboration.
• 5–6x acceleration in large model inference.
• Over 99% quantization fidelity.
• Over 50% reduction in memory usage.

Additionally, a vehicle-grade on-device AI security defense system ensures privacy protection and data compliance while balancing safety and performance.

Si Luo demonstrated several on-device intelligent scenarios, including:

• Full-vehicle, always-on wake-free interaction.
• Proactive intelligent welcome.
• Seamless smart parking payment.

Starting with dozens of proactive intelligent scenarios, AutoOmni is ushering in a new paradigm of AI-native, proactive intelligence for smart vehicles.

"Smart vehicles are the largest pilot plant and large-scale stress-testing platform for embodied intelligence," Si Luo emphasized. Banma Intelligence is now evolving from AutoOmni to EdgeOmni, expanding from single-agent intelligence to collective intelligence centered around the user, making AI terminals better understand "my world."

Disclaimer: The content above (including any images or videos) is uploaded and posted by a user of NetEase Hao, a social media platform. This platform only provides information storage services.