Banma Pushes Cars as Embodied AI Testbed at Yunqi; HarmonyOS 7 Developers Cite AI Tools

On the first day of the conference, Alibaba Group CEO Wu Yongming said machines are becoming the main force of thinking and that thinking will become a scalable commodity like power during the industrial revolution, according to Leiphone. Wu said future machine thinking volume would be 1,000 times that of humans, compared with less than 3 percent today, leaving room for at least tens of thousands of times growth.

Banma Intelligent CPO Cai Ming said cars have a more urgent demand for AI, especially agents, than phones and AI glasses, and will become the embodied intelligent terminal with the best AI experience. He described cars as high-frequency, enclosed and naturally connected to payments and ecosystem revenue sharing, making them the largest, most scalable and most mature scenario for embodied AI, a pilot production field between lab demos and mass deployment.

Zhang Yongwei, chairman of the Chebaihui Research Institute, said smart cars and embodied robots are forming a related industry cluster that shares parts of the technology base, supply chain and innovation resources, and that moving from cars to embodied intelligence has entered practical validation.

Banma said its cockpit systems have reached tens of millions of installations, its AI-in-car business covers 68 percent of mainstream automakers, its edge-intelligence market penetration is 63 percent, and it has more than 400 ecosystem partners. CTO Si Luo said edge AI has reached 80 to 90 percent of the interaction capability of cloud models with 10 times the parameters, and cloud AI will be fully driven by large models; the cloud-edge integrated cockpit architecture will become mainstream starting this year.

The company argues the business model for scaled embodied AI is an intent economy. In the internet era, services competed for search keywords; in the mobile internet era, they competed for attention in apps and feeds. In the large-model era, according to Banma, the front-end model reads user intent through context and memory and directly calls back-end service tools, including across apps and devices. Cars, because they are high-frequency, private and sensor-rich, are likely to be an early landing ground, the company said.

In Banma's proposed chain, AI identifies intent, an agent completes the task, the user pays, and carmakers and ecosystem partners share revenue. Cai compared the split to Google's Android phone system, where the operating-system licensor, hardware maker and software developer can all profit. The main threshold is whether machines can understand both the physical world, such as avoiding a dog or holding a cup, and the user's inner world, including habits, health and preferences, he said.

The supporting technology includes AutoOmni, a full-modal edge large model. Banma said AutoOmni-4B improves cockpit text understanding, visual scene understanding and voice-emotion recognition compared with leading open-source edge models. At Yunqi, it launched AutoOmni2.0-23B-A3B. The company also described an AutoOmni chip ecosystem that supports stable multi-task concurrency, speeds up large-model inference by five to six times, maintains quantization fidelity above 99 percent, halves memory use and adapts quickly to Qualcomm, Nvidia, Unisoc and Xpeng Turing AI platforms.

Banma co-CEO Hao Fei said edge models represent an interaction paradigm shift built on always-on, wake-free operation plus context and memory. A user spends about one hour a day in a car but may use wake-word AI services for only a few minutes; AutoOmni can observe, listen and remember for longer, while cross-device context brings in more of the user's life scenarios, he said.

Banma's three-layer business includes AliOS and open Android at the bottom, cockpit AI solutions such as cloud-side AutoClaw and edge-side AutoOmni in the middle, and in-car ecosystem services such as car-use and internet-service agents on top. It has more than 400 content and service providers, with over 50 signed for revenue sharing or joint operations. Alipay's AI cockpit plus AI payment has been connected, and agents such as Taobao Flash Purchase and Yuanshen short-drama are demonstrated on a Denza Z9GT at the Banma booth.

The car-side AI has already been commercialized, according to Leiphone. Banma launched Shenxingzhe 8, described as the first full-modal edge-model vehicle based on the 8397 platform, and said it ranks first in edge-model market share with design wins from 10 automakers. The company plans to extend edge AI interaction from cockpits to phones, smart homes and wearables, and to apply world models beyond driving and in-car physical understanding to robots, games and digital twins.

In the HarmonyOS 7 ecosystem, ifanr reported that three young developers, Wang Xueyuan of Pixel Artisan PRO, Meng Lingya of Xiangji and Deng Ke of Xiaoqiuquan, showed how small teams use system-level capabilities. HarmonyOS 7 offers Agentic Coding, visual AI, on-device 3DGS reconstruction and precise tap-to-share, letting developers carry ideas across devices and scenarios, according to the report.

The developers said they do not rely on vibe coding from scratch, which rarely reaches commercial completeness. Instead they embed AI in existing workflows. Pixel Artisan PRO uses AI to turn ideas into testable prototypes; Xiaoqiuquan's developer first defines goals, boundaries and test cases, then lets AI help. HarmonyOS's DevEco CLI packages Harmony tools, official knowledge and Skills into AI-callable command-line interfaces for code generation, compilation, deployment and debugging.

Specific kits solve concrete problems. Xiangji uses FAST Kit's natural-language sorting so version 2 appears before version 10. Pixel Artisan PRO uses Spatial Recon Kit to reconstruct a fine 3D model on device after a user walks around an object. Xiaoqiuquan uses a system-level A2A interaction framework so users can ask Xiaoyi for recent tennis activity without opening the app, subject to app permissions and system rules.

New Huawei device shapes are also changing app design. The wide foldable Pura X Max, wide bar Pura X View and tri-fold Mate XT 2, released in September, offer different screen proportions. Pixel Artisan PRO said a wider screen separates canvas, palette, toolbar and reference images; Deng Ke prefers Pura X Max's 1.414-to-1 ratio for video and motion analysis. Nameo uses tap-to-share so material moves between phone, tablet and PC.

HarmonyOS adaptive layout and the updated Parallel Vision help apps handle unfolding, rotation, split screen and keyboard changes without rebuilding controls for each device. The developers said AI and system capabilities make the same tasks easier, but they also make previously shelved ideas worth trying, leaving them busier but with more freedom.