Alibaba's Banma ships AutoOmni 2.0: 3B-active edge model nears 10x-larger cloud models

机器之心 · wechat · 2026-09-24

At the Yunqi Conference, Alibaba's Banma unveiled AutoOmni 2.0-23B-A3B, a 23B-total / 3B-active MoE edge model for smart cockpits. The company claims it matches a 10x-larger cloud model on ordinary cockpit tasks and reaches 80–90% on complex ones. CTO Si Luo admitted that just three weeks earlier he judged a demo impossible within the quarter — edge model progress is outpacing even practitioners' own expectations.

The article frames the second phase of AI-in-cars: hardware is ubiquitous (70.5% L2 penetration in H1 2026) yet paid adoption of higher-level ADAS is below 10% — users value the capability but expect it free. Edge deployment faces three walls (memory, latency, concurrency) plus privacy and cost constraints; Banma's answer combines sparse MoE with quantized deployment (5–6x faster inference, 99%+ quantization fidelity, 50% less memory), a brain/cerebellum split between the AutoOmni model and agent execution framework, SQLite+RAG on-device memory, rejection mechanisms, and a security stack with 98%+ recall against jailbreak attacks.

The strategic read: the battleground is shifting from cloud model capability to the edge; Banma has adapted to ten chip vendors and plans to extend the same execution framework into assisted driving toward cockpit-driving fusion.

Original post →

More from Embodied

Embodied channel →