At the two Arm Create events held in Shanghai and Shenzhen, developers deeply explored key decisions across infrastructure, model selection, deployment, optimization, and validation, jointly seeking the best path to transform AI models into real-world software applications.
A high-performance model is just the beginning. Building artificial intelligence (AI) applications also requires making a series of key decisions around model architecture. When factors such as latency, cost, privacy, power consumption, and deployment scenarios begin to influence design, these decisions often determine the application's final performance.
At the Arm Create Shanghai and Shenzhen events in 2026, developers explored these related design decisions across cloud, edge, mobile, graphics processing, and physical AI by attending keynote speeches, partner shares, deep-dive technical sessions, and participating in hands-on practice.
Image: Shantu Roy, VP of Developer Relations at Arm, delivering a keynote speech at the Arm Create event
Arm's latest series of hardware and software achievements were also a core topic of discussion. The Arm AGI CPU and the new Neoverse CSS N4 provide more options for cloud infrastructure; Arm's second-generation Mobile Compute Subsystem (CSS for Mobile 2) brings a new on-device AI experience to mobile devices; and the Arm Mali G2-Ultra NX introduces neural processing capabilities to mobile platforms. At the same time, the newly launched Arm AI Portal helps developers discover models that have been verified and optimized for Arm platforms, understand their compatibility with target platforms, and access relevant code and deployment workflows.
However, for developers, the key is not what features each platform or tool has, but whether they can help developers make decisions, conduct experiments and tests, and continuously optimize software.
In the keynote speech at Arm Create Shanghai, Arm raised four key questions regarding this industry challenge:
· Which intelligence capability fits the corresponding task?
· Where should the workload be deployed?
· How do components achieve system-level synergy?
· Can the application achieve expected performance on real target hardware?
Subsequently, through technical sharing and hands-on labs, five key insights for developers building and deploying AI were further summarized.
1. Determine the infrastructure required for agent applications
When applications have agent capabilities, the model is only part of the challenge. Agents can dynamically generate code, call tools, and execute actions, thus placing new demands on underlying infrastructure.
At the Arm Create Shanghai event, Tencent Cloud's open-source AI Agent security sandbox Cube Sandbox presented real-world case studies. Discussions centered on key issues such as isolation, concurrency, security, state management, resource management, and cost, exploring how to reasonably allocate different workloads at the platform level, allowing developers to focus more energy on application innovation itself.
Image: Tencent Cloud shared the open-source AI Agent security sandbox Cube Sandbox and infrastructure solutions supporting agent applications at Arm Create Shanghai.
The insight for developers is that agent applications need to build a controlled execution environment beyond the model, enabling consistent management of tool calls, operation execution, state information, and computing resources as the application scales.
2. Match workloads and choose the right model
Model selection should start with the workload, not just model size. Instant response features in mobile apps, agents capable of cloud inference on massive documents, and edge vision workloads may all lead model selection in vastly different directions.
At the Arm Create Shenzhen event, guests from Alibaba Tongyi Qianwen, ModelBest, Tencent Hunyuan, and Ultralytics participated in a panel discussion. The focus was no longer limited to comparing models but centered on application suitability: workload requirements, deployment locations, and actual performance after deployment. The core principle is clear: match the model to the task, and run the task in the most suitable location.
The Arm AI Portal is built around this need. It helps developers discover models verified and optimized for Arm platforms, evaluate model compatibility with target hardware, and access code and deployment information, thereby shortening the cycle from model evaluation to actual implementation.
Image: Developers can use the Arm AI Portal to evaluate model compatibility on specific Arm platforms based on performance data, providing reference for model selection and deployment decisions.
3. Evaluate and optimize the entire application
When AI applications suffer from response delays or poor efficiency, the model is not the only area to investigate. Runtime environments, data transmission, preprocessing pipelines, application logic, or underlying computation also consume time and resources. For example, the performance bottleneck of a voice assistant might lie in audio preprocessing; when mobile AI features run continuously, they may hit memory limits or thermal constraints.
Therefore, system-wide performance analysis and optimization have become part of the development process. Arm provides various tools and technologies to support this process: Arm Performix and Arm Performance Studio help developers understand where workloads consume time and resources; the second-generation Arm Scalable Matrix Extension (SME2) and Neural Graphics tools offer different ways to enhance performance on Arm-based platforms. Truly effective optimization should be based on data obtained from workload analysis and revolve around the most critical constraints, including response latency, power consumption, cost, and sustained performance.
For cloud workloads, optimization might involve identifying CPU or memory hotspots to improve the efficiency of existing computing resources; for mobile AI, it means balancing response speed and power consumption; and for graphics workloads, it requires maintaining image quality while ensuring smooth frame rates and thermal headroom.
4. Determine the optimal execution location for each workload
Not all AI application components should be deployed in the same location. Developers need to determine: which workloads should run locally on the device; which are suitable for nearby edge infrastructure; and which require the scalability and shared services of the cloud.
Computing power determines where a workload "can" run, while product requirements determine where it "should" run. Latency, privacy protection, offline capability, local control, cost, and available compute all influence this deployment choice. A single application can split tasks across multiple locations. The goal is not to find a single deployment location for the entire application, but to place each functional module in the most suitable position based on its requirements.
Arm computing platforms cover various deployment scenarios including cloud, edge, mobile devices, and graphics computing, while the Arm AI Portal helps developers link model selection with specific Arm target platforms and deployment paths.
5. Maintain workflow portability and validate on real target platforms
AI applications rarely stay in the same environment throughout their lifecycle. Teams may prototype locally, test in the cloud, migrate some workloads to the device side, and then optimize for mobile, edge, graphics, or physical AI targets. If models, runtimes, toolchains, and performance workflows need to be completely replaced each time, this migration creates more workload.
During the dialogue session at Arm Create Shanghai, Eric Ernst from the Aston Martin Aramco F1 Team and Chen Xiaoli from Unitree Robotics used F1 cars and humanoid robots as examples to discuss an engineering challenge: software, hardware, data, sensors, and control systems must work together in complex real-world physical environments, where actual conditions are often unpredictable.
Image: Arm, Aston Martin Aramco F1 Team, and Unitree Robotics jointly explore how software and hardware collaborate in complex physical scenarios.
This also makes portability and verification closely related issues. If models, runtime environments, toolchains, and performance workflows are constantly changing, every migration between cloud, edge, mobile, graphics, and physical AI introduces additional complexity. Therefore, developers need to adopt tools, models, libraries, and workflows that remain consistent as applications migrate. Meanwhile, no matter how thorough the preliminary simulation and testing are, final validation must still be completed in the target environment where the product actually runs. Arm Create's hands-on course on Physical AI further extends this concept, covering practices such as simulation-to-robot workflows, application deployment, trust mechanisms, telemetry capabilities, and device discovery.
In this process, SME2, KleidiAI, Arm Performix, Arm Model Context Protocol (MCP) servers, Neural Graphics tools, Arm Device Connect, and Arm AI Portal support different stages of the workflow, helping developers complete key tasks from model discovery, workload optimization, and performance evaluation to connecting intelligent software with devices.
Image: At Arm Create, Arm Device Connect demonstrated the open layer for connecting AI agents with physical devices.
Ultimately, the core practical experience is: consider portability early; keep workflows as consistent as possible when deployment choices change; and validate the complete system on the final target platform.
Building for the next generation of innovation
Demonstrating practical methods and best practices for AI development is the essence of Arm Create events. In both Shanghai and Shenzhen, developers moved from concepts in keynote speeches to deeper technical sessions, live demos, hands-on courses, and direct exchanges with Arm engineers and ecosystem partners. This immersive practical experience also provided a clear path for developers' subsequent work. In the Arm Create Shanghai survey, 97% of respondents stated that the event deepened their understanding of how to build, optimize, and deploy AI applications on Arm platforms.
Image: Demo of code optimized for Arm platforms at Arm Create attracts developers
As AI applications cover an increasingly wide range of target scenarios, developers are no longer focused solely on optimizing for a single platform, but on ensuring continuity of code, tools, skills, and performance workflows across different deployment scenarios. Developers need a workflow that migrates with the application, rather than starting from scratch every time the deployment scenario changes. The unified Arm architecture and software foundation are helping developers extend their existing achievements to various application deployment scenarios, including cloud, edge, mobile devices, graphics computing, and physical AI.