Arm has introduced a new platform called Arm AI Portal. The tool is designed to help developers build AI applications faster on Arm-based hardware.
The company announced the launch through a post from Sharbani Roy, VP of AI and Developer Platforms at Arm. The portal is meant to simplify a process that has traditionally taken weeks of searching, testing and tuning.
Arm's compute platform is used across cloud computing, edge devices and physical AI systems like robots. More than 22 million developers currently build software for these systems.
What the AI Portal Offers
The AI Portal gives developers a single place to find AI models that are already optimized for Arm hardware. Users can compare details like latency, memory use and file size before choosing a model.
Code examples and deployment workflows are also included on the platform. This is meant to help developers move from testing to launch more quickly.
At launch, the portal includes pre-optimized models such as Alibaba's Qwen, Google's Gemma and Ultralytics' YOLO. These models work with runtimes including ExecuTorch, LiteRT and ONNX-RT.
Arm is working with partners including Alibaba, Raspberry Pi and Ultralytics on this effort. The models are also accessible through Hugging Face, a popular platform for sharing AI models.
Coding agents can access AI Portal resources directly through a system called MCP. This is meant to support the shift toward agentic AI, where software agents help write and manage code.
Arm said it will soon add tools that let developers bring their own models to the platform. This would include private or proprietary models, which developers could then optimize for Arm hardware.
These tools are currently available through early access. A general release is expected at a later date.
Performance Gains From Early Testing
Arm shared some early results from models optimized through the new platform. A speech model called Qwen3-TTS ran more than four times faster on a vivo X300 smartphone compared to before optimization.
That speedup came from using single-thread execution and mixed quantization. It also relied on a chip feature called Scalable Matrix Extension-2, or SME2.
A vision model called Ultralytics YOLO26n also showed gains. Performance improved by more than 40% on the same smartphone using a mix of FP16 and FP32 processing with SME2.
On a Raspberry Pi 5, the YOLO26n model showed similar gains using FP16 and INT8 processing. These tests used the device's built-in NEON technology.
Arm said the AI Portal connects several of its chip technologies to optimized software. This includes SVE and SME, along with neural accelerators built into some Arm chips.
One example is Arm's CSS for Mobile 2 platform. That platform includes models accelerated by SME2 and GPUs with neural accelerators, both of which are available through AI Portal.
Arm has invested in software tools for developers for many years. The company said AI Portal is the next step in that effort, aimed at supporting AI development across its full range of hardware.
The platform is now live for developers, with expanded tools for custom models still in early access ahead of a wider rollout.
