Arm Unveils 'AI Portal' as AI Agents Tap Arm Optimization

Arm has unveiled the "Arm AI Portal" to support AI application development. Its scope covers Arm computing platforms across cloud, edge, and physical AI.

As agentic AI expands beyond the cloud into edge and physical AI, the complexity of AI application development has also increased. As a result, developers are now tasked with finding suitable models across different model, runtime, and hardware environments, comparing performance, and then optimizing them.

Arm created the Arm AI Portal to ease this development burden. The portal lets users explore preoptimized models that include performance and accuracy data, and compare latency, memory usage, and model size.

The Arm AI Portal also provides code examples and deployment workflows. It supports language, speech, vision, and neural graphics, and at launch it offers models from Alibaba Qwen, Google Gemma, and Ultralytics YOLO.

The target models run using ExecuTorch, LiteRT, and ONNX-RT as their runtimes. They are based on support from ecosystem partners at Alibaba, Raspberry Pi, and Ultralytics. The optimized models provided by the Arm AI Portal also offer examples of performance gains in real Arm-based environments.

Developers can use Arm-optimized models through Hugging Face. Coding agents can access AI Portal resources through MCP. This allows them to use Arm-optimized AI models and related resources while keeping their existing development environments.

Arm plans to bring in-house models and proprietary models to the AI Portal in the future. It also plans to provide tools that can perform performance analysis and optimization in Arm-based environments.

Agent-ready AI resources can be accessed via early access before their full release.

As an example of performance gains from individual AI models using Arm technology, Qwen3-TTS was cited first. On the vivo X300 smartphone, Qwen3-TTS delivered more than a 4x speedup by using single-threaded execution and mixed quantization. In this process, a Q8_0 token and code predictor was used, and Arm SME2 (Scalable Matrix Extension 2) served as the acceleration method.

Ultralytics YOLO26n was also presented alongside use cases on the vivo X300 smartphone and the Raspberry Pi 5. On the two devices, Ultralytics YOLO26n delivered more than 40% performance gains after applying Arm technology. The YOLO26n optimization on the vivo X300 used SME2 and FP16-based single-threaded execution, while the YOLO26n optimization on the Raspberry Pi 5 used NEON, FP16, and INT8 mixed quantization.

The scope of the Arm AI Portal is not limited to specific devices. Its target covers Arm computing platforms across cloud, edge, and physical AI. Developers can explore AI models for each target platform and use optimized software. Examples include vision models for robotics, generative AI for smartphones, and LLMs for cloud CPU workloads.

Arm plans to connect Arm computing technologies such as SVE, SME, and neural accelerators with optimized software through the AI Portal to improve access to development resources. A representative example is Arm CSS for Mobile 2. Through the AI Portal, developers can use SME2-accelerated models and GPU-related resources equipped with neural accelerators.

Arm plans to expand the reach of its existing software ecosystem into AI development through the AI Portal. The portal is built on the software ecosystem Arm has established over decades. Through this, Arm aims to support more than 22 million developers and agents, and make it easier to find and use AI software optimized for target hardware.

Arm is pursuing a strategy that allows not only developers but also AI agents to discover machine-discoverable models, performance data, and workflows. The goal is to provide a development environment suited to the era of agentic AI.

Source: TECHWORLD · Park Gyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407389

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407389