Key takeaway
- How Arm processors speed up AI vision on phones, what changed since 2025, where to run vision workloads, and a sourced 3 to 5 year outlook for teams.
Arm processors run AI vision on mobile devices by combining CPU cores with matrix extensions (SME2), GPUs and SoC-level accelerators, and by using shared software libraries that speed up common vision and machine learning kernels. For an enterprise, the practical question is no longer whether a phone or tablet can run a vision model, but which workloads should run on the device, which should stay in the cloud, and who is accountable for what the camera sees.
This guide covers what Arm's current mobile platform offers, what has changed since 2025, how to decide where a vision workload runs, the governance and cost points that matter, and what is likely to change over the next three to five years. For the wider picture, see our guides to computer vision applications with agentic AI and generative AI in computer vision.
What do Arm processors contribute to AI vision on mobile devices?
Arm designs processor architectures and licenses them to chip makers, so the AI vision capability of a given phone depends on the SoC vendor's choices as well as on Arm's designs. Arm's own announcement of the Lumex compute subsystem on 10 September 2025 describes real-time, private, on-device AI as the new expectation, and names Apple, Samsung and MediaTek among the companies integrating AI acceleration.
| Component | What it does for vision | When it fits | Current status (Arm sources) |
|---|---|---|---|
| CPU with SME2 | Accelerates the matrix maths behind CNNs, transformers and LLMs directly on the CPU | Mixed workloads, small and medium models, no extra accelerator to manage | Armv9.3 feature; Arm cites up to 5x AI performance for the SME2-enabled C1 CPU cluster in Lumex |
| GPU (Mali-G1-Ultra) | Graphics and parallel inference | Camera effects, AR rendering alongside inference | Arm cites AI inference up to 20% faster than the previous generation |
| Ethos-U85 NPU | Embedded neural acceleration beside Cortex-M or Cortex-A cores | Cameras, sensors and robots with tight power budgets | 256 GOP/s to 4 TOPS at 1 GHz, native transformer support, 20% more energy efficient than Ethos-U55 and U65 |
| Kleidi libraries | Optimised kernels for ML and computer vision frameworks | Teams using LiteRT, ExecuTorch, MediaPipe, ONNX Runtime, MNN or OpenCV | KleidiCV is integrated in OpenCV 4.11; Arm cites up to 4x for blur, filter, rotation and resizing |
One practical example from Arm: neural camera denoising runs at over 120 frames per second at 1080p, or 30 frames per second at 4K, on a single CPU core of the new platform. These are vendor figures, so test them on your own devices and models before they enter a business case. For heavier vision on the factory floor, our guide to computer vision for automated assembly line inspection shows how the same ideas apply beyond phones.
What has changed in Arm-based mobile AI vision since 2025?
- Matrix acceleration moved onto the CPU. SME2 is available on iPhone 16 and Apple M4-series chips and on Android devices based on Lumex C1-series CPUs, according to Arm's developer pages.
- Frameworks pick up the gains. Arm says KleidiAI is integrated into ExecuTorch, LiteRT, MNN and ONNX Runtime, so applications benefit without code changes, although some published results depend on quantised models or specific settings.
- Android's accelerator API is on its way out. Google's documentation states that NNAPI was deprecated in Android 15 and that it expects the majority of devices in future to use the CPU backend. It recommends alternatives such as the TF Lite GPU runtime for performance-critical work. Plans that depended on vendor NNAPI drivers need review.
- Vision transformers and on-device generative models are now in scope. Arm's developer pages cover Vision Transformer inference on Android with KleidiAI and MNN, and report up to 3.9x faster inference with SME2 for the SqueezeSAM model behind Instagram's Cutouts tool.
How do you decide where a mobile vision workload should run?
| Workload | Best fit | Why |
|---|---|---|
| Face unlock, document capture, camera effects | On the device | Latency and privacy; Arm cites both, with cost, as drawbacks of cloud-first AI |
| Augmented reality overlays | On the device, CPU and GPU together | Frame-rate sensitive; see our AR and VR guide |
| Large or frequently retrained models | Cloud or edge server | Model updates and memory limits are easier to manage centrally |
| Fleets of cameras, drones or robots | Edge platform with on-device agents | Needs central policy, monitoring and updates across many devices |
For the last row, NexaStack is an agentic operating platform for physical AI that runs across on-premises, cloud and hybrid environments, and its own site lists on-device intelligence among its capabilities. Connectivity matters too: our note on AI in 5G networks covers what faster links add to edge vision.

What are the governance, security and cost points for on-device vision?
Processing images on the device keeps raw frames off the network, which Arm positions as a privacy and cost advantage over cloud-first AI. It does not remove legal duties. Under the EU AI Act, the prohibitions on practices such as untargeted scraping of facial images and emotion recognition at work and in education have applied since 2 February 2025. According to the European Commission, high-risk obligations for use cases such as biometrics apply from 2 December 2027, after the AI Omnibus changes, and the transparency rules apply from August 2026.
Treat each vision feature as a decision with an owner, a policy and an evidence trail. The Compliance Agent enforces policy, approvals and evidence so AI execution stays audit-ready, according to its own page. On cost, the main levers are model size, quantisation and how often a task needs a cloud round trip.
What is a sensible adoption sequence?
- List the vision tasks and mark each as on-device, edge or cloud, using the table above.
- Benchmark two or three target devices with your own models, using the frameworks you already ship.
- Check the runtime path on Android, given the NNAPI deprecation.
- Apply the privacy, biometric and AI Act review before the pilot, not after it.
- Roll out in stages and monitor accuracy, battery and thermal behaviour per device class.
When should you not rely on on-device Arm vision?
Arm's published speed-ups come from vendor tests under varying conditions, and Arm's own page does not state the framework versions behind each figure. Do not commit to a performance target from a headline number. Avoid on-device deployment where models change weekly, where devices are too varied to test, or where a regulator expects centrally held evidence of every decision. In those cases, run inference at the edge or in the cloud and keep the device for capture and pre-processing.
What will change in Arm-based mobile AI vision over the next 3 to 5 years?
The sourced signals below point towards matrix acceleration becoming standard on the CPU, more AI-capable phones, and a firmer regulatory frame around biometric vision. Forecasts are the publishers' own and can change.
| Expected change | Horizon | Source and caveat |
|---|---|---|
| SME and SME2 add over 10 billion TOPS across more than 3 billion devices | By 2030 | Arm projection (Arm Newsroom, 10 September 2025); a vendor forecast |
| GenAI smartphone shipments reach 912 million units, a 78.4% CAGR from 2024 | 2028 | IDC, 30 July 2024. IDC defines these as phones with an NPU of 30 TOPS or more (int-8); the forecast predates later market data |
| CPU-first inference on Android as NNAPI is phased out | Ongoing from Android 15 | Google Android documentation; Google states its expectation, not a date |
| High-risk biometric obligations apply under the EU AI Act | 2 December 2027 | European Commission, after the AI Omnibus |
| Ethos-class NPUs carry transformer models in cameras, sensors and robots | Current direction | Arm product page for Ethos-U85 (native transformer support); adoption timing is not stated |
The planning implication is that device capability will stop being the main constraint. Model governance, update paths and evidence will be. Teams that build a vision decision layer now will adapt more easily as the hardware improves.
To map one on-device vision decision your agents should make, and the context and policy it depends on, book a Context Workshop. See what happens in the workshop.
What should you plan for in 2027?
- 2 December 2027: the European Commission states that high risk rules for use cases in sensitive areas, including biometrics, apply from this date after the AI Omnibus. See the Commission AI Act page. Review any face or biometric feature before its pilot.
- 2 August 2028: the same page gives an extended transition period to this date for high risk systems in regulated products, which matters if a device feature ships inside one.
- Benchmark on your own target devices, because vendor speed ups vary with framework and settings.
Frequently asked questions
Which Arm technology speeds up AI vision on phones?
Arm names the SME2 CPU extension, the Mali GPU line and the Kleidi software libraries. Arm cites up to 5x AI performance for the SME2-enabled C1 CPU cluster in its Lumex platform.
Does the Ethos-U85 run on phones?
Arm positions the Ethos-U85 for embedded and IoT devices, from 256 GOP/s to 4 TOPS at 1 GHz, with native transformer support. Phone SoCs use their vendors' own accelerators or the CPU and GPU.
Is Android NNAPI still the way to run vision models?
No. Google's documentation states that NNAPI was deprecated in Android 15 and recommends alternatives such as the TF Lite GPU runtime for performance-critical workloads.
Do EU rules affect facial recognition on mobile devices?
Yes. According to the European Commission, certain biometric practices have been prohibited since 2 February 2025, and high-risk biometric use cases follow from 2 December 2027.
What will change in mobile AI vision over the next 3 to 5 years?
Arm projects SME and SME2 will reach more than 3 billion devices by 2030, and IDC forecast 912 million GenAI smartphones shipping in 2028. Expect more on-device vision and tighter governance of biometric use.
Can I trust vendor speed-up figures?
Treat them as indicative. Arm's own pages do not state the framework versions behind every figure, so benchmark on your target devices.
Next Steps in Advancing ARM-Based AI Vision
Talk to our experts about implementing AI vision solutions. Discover how industries and various sectors leverage Edge AI and real-time visual processing to enhance decision-making with autonomous AI workflows. Harness AI-powered automation to optimize computer vision tasks, improving efficiency, accuracy, and responsiveness.
Follow XenonStack on LinkedIn: XenonStack on LinkedIn for new guides on agentic AI, data and platform engineering.
Ready to put AI agents to work? Explore ElixirClaw, the agentic OS platform to build, orchestrate and govern AI agents.
Explore the XenonStack platform family: ElixirData · ElixirClaw · ElixirHub · NexaStack · MetaSecure · Neural AI · AgentGRC · AgentSRE · AgentAnalyst · AgentInstruct · AgentRAI · AgentScrum · AgentSearch · AgentForce · AI Aviator · SAIF Aviator · CohortIQ · XenonStack AI · XenonStack UAE