Arm Launches AI Portal for Optimized Models Across Devices

Arm has introduced AI Portal, a centralized platform designed to help developers and AI coding agents discover, optimize, and deploy machine learning models across the company's compute ecosystem.

The platform addresses a growing challenge as AI workloads expand beyond cloud infrastructure into edge devices and robotics. Developers building across different hardware targets—from smartphones to cloud servers—currently face weeks of benchmarking and optimization work before deploying models. AI Portal aims to compress that timeline by providing pre-optimized models with transparent performance data.

Pre-optimized models with documented performance

At launch, AI Portal includes models from Alibaba Qwen, Google Gemma, and Ultralytics YOLO, running on frameworks including ExecuTorch, LiteRT, and ONNX Runtime. The platform covers language, speech, vision, and neural graphics workloads across Arm-based hardware.

Early performance benchmarks demonstrate substantial gains from Arm-specific optimizations. Qwen3-TTS achieved over 4x speedup on a vivo X300 smartphone using single-thread execution and mixed quantization, accelerated by Scalable Matrix Extension-2 (SME2). Ultralytics YOLO26n showed over 40% improvement using FP16 versus FP32 on the same device with SME2, and similar gains with mixed quantization on Raspberry Pi 5 using NEON instructions.

Developers can compare models by latency, memory footprint, and size, with access to code examples and deployment workflows. The platform will soon support custom model uploads, including proprietary models, for performance analysis and optimization on Arm hardware.

Agent-ready AI resources

As software development increasingly incorporates AI coding assistants, AI Portal makes its resources machine-discoverable. Arm-optimized models are available through Hugging Face, while Portal resources are accessible to coding agents through the Model Context Protocol (MCP). This agent-ready approach is currently in early access ahead of general release.

The platform connects Arm's hardware technologies—including Scalable Vector Extension (SVE), Scalable Matrix Extension, and neural accelerators—with optimized software implementations. Arm CSS for Mobile 2, for example, includes models accelerated by SME2 and GPUs with neural accelerators available through the portal.

Why it matters

The fragmentation of AI deployment across cloud, edge, and embedded systems creates significant friction for developers. With AI workloads moving to smartphones, IoT devices, and robotics, optimizing models for specific hardware architectures has become a bottleneck. Arm's unified portal approach could reduce time-to-deployment while improving performance, particularly important as the company's architecture powers the majority of mobile devices and gains traction in cloud data centers. The platform's agent-ready design also anticipates a development workflow where AI assistants handle more optimization tasks.

Arm reported these details in a company announcement. The platform is available now, with ecosystem support from partners including Alibaba, Raspberry Pi, and Ultralytics.