Arm has launched a software catalog for finding and deploying AI models tuned to its processors, giving developers and coding agents one place to compare model performance and access deployment material. Arm announced the Arm AI Portal on September 8, describing it as a way to discover, optimize, and deploy AI software across cloud, edge, and physical systems.
Arm says the portal connects a developer community of more than 22 million people and their agents to optimized software. The initial service lists task-specific models with performance and accuracy data, comparisons for latency, memory, and size, plus code examples and deployment workflows. Arm says tools for analyzing and optimizing a developer’s own model, including proprietary models, are planned for a later release.
A catalog built for deployment, not just discovery
The launch covers language, speech, vision, and neural graphics workloads. Arm names Alibaba Qwen, Google Gemma, and Ultralytics YOLO among the first model families, with runtimes including ExecuTorch, LiteRT, and ONNX Runtime. Alibaba, Raspberry Pi, and Ultralytics are listed as ecosystem partners. The portal’s stated job is to narrow the gap between choosing a model and getting it onto a particular Arm target.
Arm is also publishing early performance examples, and the numbers need to be read as vendor-reported results rather than universal guarantees. The company says Qwen3-TTS achieved more than a four-times speedup on a vivo X300 smartphone using single-thread execution, mixed quantization, and Arm SME2 acceleration. It also reports more than 40 percent improvement for Ultralytics YOLO26n in comparisons using FP16 and INT8 configurations on a vivo X300 and Raspberry Pi 5.
Those details matter because an AI model’s headline size says little about how it behaves on a real device. The runtime, quantization format, memory path, thread count, and accelerator all change the result. Our NPU explainer covers why dedicated neural hardware is useful, while our local AI hardware analysis looks at the tradeoffs that remain when workloads move out of the cloud.
Arm is positioning the portal across its full compute range instead of treating smartphone, Raspberry Pi, robotics, and cloud CPU deployment as separate software worlds. That could be useful for teams that need to move the same application between different Arm devices, although an optimized model listing is not the same as proof that every model will run identically across those targets.
The portal is designed to be machine-discoverable as well as human-readable. Arm says resources can meet developers and agents where they build, and early-access tools are available ahead of general release. That approach fits the way coding agents increasingly select libraries, models, and runtime settings, but it also raises the importance of clear test conditions and versioned performance data.
Arm has not presented the portal as a new processor or a replacement for existing model hubs. It is a software and workflow layer around Arm’s existing CPU, GPU, and neural acceleration technologies. The next useful test will be whether developers can reproduce the published gains on their own hardware and whether the catalog expands beyond the launch models. For now, Arm is making optimization evidence easier to find before a model reaches the device.
