AI PORTAL
  • Models
  • Code
  • Tools

About the models on this Arm AI Portal

The Arm AI Portal (the “Portal”) provides details of AI models developed, optimized or tested for use on Arm-based platforms. Models listed on the Portal may have been developed by Arm or by third parties. Unless expressly stated otherwise, Arm did not develop or train the underlying third-party model and is not responsible for its original intended behavior or use.

Information accompanies each model listing and describes the model’s provenance, together with relevant information about the original third-party model, its developer and repository, where applicable. Inclusion on the Portal of a model developed by a third party or based on a third-party model does not constitute Arm’s endorsement or certification of that third-party model or any third-party work relating to it.

Except as expressly stated in the information accompanying a specific model, Arm makes no representation as to the accuracy, safety, security, non-infringement, legal compliance, suitability for production use or fitness for any particular purpose of any model, converted or optimized version, or related outputs.

Inclusion of a model on the Portal does not itself grant you any rights to use that model. Use of each model is subject to the applicable license terms, usage restrictions and documentation. You are responsible for reviewing that information and independently evaluating the model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements. Any use of or reliance on a model, its outputs or related materials is at your own risk. Where the Portal links to repositories or other materials controlled by third parties, those repositories and materials are outside Arm’s control and may change without notice.

Use of the Portal is subject to the Arm Website Terms and Conditions of Use.

Models

Filters

1

Device Class

Edge LinuxEthos-U NPUMobile CPUCloud CPU

Evaluation Target

Raspberry Pi 5vivo X300Alif DK-E8AWS Graviton G4

Arm Technology

SME2

Task

Text GenerationImage ClassificationObject DetectionFeature ExtractionImage-to-ImageAutomatic Speech RecognitionImage SegmentationDepth EstimationText-to-SpeechAudio ClassificationKeypoint DetectionImage-Text-to-TextText-to-ImageTranslationZero-Shot Image Classification

Vendor

Runtime

Quantization

Format

Execution Backend

Model Size

≤88.6 MB

Average Memory

≤190 MB

Peak Memory

≤190 MB
    48 results

    ResNet-50 INT8

    Raspberry Pi 5

    INT8-quantized 50-layer residual network for 1000-class image classification. It uses symmetric per-channel quantization with XNNPACK support, accepts 224×224 RGB inputs with ImageNet normalization, and returns raw logits for class prediction.

    image-classification25.9 MB
    ExecuTorch

    GoogLeNet INT8

    AWS Graviton G4

    An INT8-quantized GoogLeNet image classification model based on an Inception-style convolutional architecture. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification7.1 MBExecuTorch

    ResNet-18 INT8

    Raspberry Pi 5

    INT8-quantized 18-layer residual network for 1000-class image classification, optimized for efficient ARM edge inference with ExecuTorch and the XNNPACK backend. It uses 224×224 RGB inputs with ImageNet normalization and returns logits for top-k prediction, offering a much smaller, faster, and lower-memory alternative to the FP32 baseline while retaining comparable classification quality.

    image-classification11.8 MBExecuTorch

    GoogLeNet INT8

    Alif DK-E8

    An INT8-quantized GoogLeNet image classification model based on an Inception-style convolutional architecture. Exported as an ExecuTorch program using the Ethos-U delegate for inference on Arm Ethos-U NPUs.

    image-classification5.5 MBExecuTorch

    ResNet-50 INT8 (per-tensor)

    Alif DK-E8

    A 50-layer residual convolutional neural network for 1000-class image classification, using INT8 per-tensor quantization to reduce size and computation while preserving strong accuracy. It accepts 224×224 RGB inputs with ImageNet-style preprocessing and produces raw logits over 1,000 categories.

    image-classification15.1 MBExecuTorch

    Wide ResNet-50 INT8

    vivo X300

    An INT8-quantized Wide ResNet-50 image classification model using a widened residual convolutional network for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification50.6 MBExecuTorch

    ResNet-50 INT8 (per-channel)

    Alif DK-E8

    An INT8 per-channel quantized 50-layer residual network for 1000-class image classification, using residual connections to support deep feature extraction while reducing model size and compute requirements. It accepts 224×224 RGB inputs with ImageNet-style normalization and outputs raw logits for class prediction.

    image-classification22.8 MBExecuTorch

    Inception v3 INT8

    vivo X300

    An INT8-quantized Inception v3 convolutional image classification model using factorized convolutions for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification25.6 MBExecuTorch

    Inception v3 INT8

    Raspberry Pi 5

    An INT8-quantized Inception v3 convolutional image classification model using factorized convolutions for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification25.4 MBExecuTorch

    Swin-T INT8

    vivo X300

    An INT8-quantized Swin-T hierarchical Vision Transformer for image classification, using shifted-window attention for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification32.5 MBExecuTorch

    DeiT-Tiny INT8

    Alif DK-E8

    An INT8-quantized tiny Vision Transformer for 1,000-class image classification, using SmoothQuant to reduce model size while preserving accuracy. It processes 224×224 RGB images with ImageNet-style normalization and outputs class logits for top-k prediction, making it suitable for efficient edge AI vision workloads.

    image-classification4.9 MBExecuTorch

    SqueezeNet 1.0 INT8

    AWS Graviton G4

    An INT8-quantized SqueezeNet 1.0 compact convolutional image classification model using fire modules for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification1.9 MBExecuTorch

    MobileNetV3-Small INT8

    Raspberry Pi 5

    A compact INT8 image-classification CNN using inverted residual blocks, Squeeze-and-Excitation modules, and hard-swish activations to deliver efficient 1,000-class visual recognition with low compute and memory requirements.

    image-classification2.8 MBExecuTorch

    Inception V3 INT8

    AWS Graviton G4

    This is an INT8-quantized version of Inception V3 optimized for edge deployment on ARM devices using ONNX Runtime, profiled on AWS Graviton G4. The model was quantized using PT2E static post-training quantization and exported to the .onnx format via the shared PT2E-to-ONNX pipeline for efficient inference on ARM Neoverse-V2 processors.

    image-classification24.9 MBONNX

    EfficientNet-B4 INT8

    Raspberry Pi 5

    An INT8-quantized EfficientNet-B4 convolutional image classification model for visual recognition. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification22.8 MBExecuTorch

    SqueezeNet 1.1 INT8

    Raspberry Pi 5

    An INT8-quantized, fire-module CNN for efficient 1000-class image classification, using per-channel weights and per-tensor activations. It provides a compact, low-latency alternative to its FP32 counterpart with minimal accuracy impact, fixed 224×224 image inputs, and standard softmax-based top-class prediction outputs.

    image-classification1.3 MBExecuTorch

    MobileNetV3-Large INT8

    Alif DK-E8

    An INT8-quantized image classification model based on a depthwise-separable convolutional architecture with squeeze-and-excitation blocks, designed for efficient 1000-class recognition. It uses per-channel quantized weights, 224×224 RGB inputs with ImageNet normalization, and outputs raw logits for class prediction.

    image-classification5.4 MBExecuTorch

    MobileNetV3-Small INT8

    vivo X300SME2

    Lightweight INT8 image classifier for 224×224 RGB inputs, producing logits for 1,000 ImageNet-style classes. Built for mobile and edge use on ARM CPUs, it combines a compact inverted-residual CNN architecture with quantized inference to reduce size, latency, and memory while preserving strong classification accuracy.

    image-classification3 MBLiteRT

    SqueezeNet 1.0 INT8

    Alif DK-E8

    An INT8-quantized SqueezeNet 1.0 compact convolutional image classification model using fire modules for visual feature extraction. Exported as an ExecuTorch program using the Ethos-U delegate for inference on Arm Ethos-U NPUs.

    image-classification1.3 MBExecuTorch

    GoogLeNet INT8

    AWS Graviton G4

    INT8-quantized GoogLeNet image classification model optimized for ARM CPUs, exported to ONNX Runtime with MLAS/KleidiAI kernels, profiled on AWS Graviton G4 (c8g.2xlarge). Accepts 224x224 RGB inputs with standard ImageNet preprocessing and outputs raw logits for 1,000 classes.

    image-classification7.3 MBONNX

    OSNet x0.25 INT8

    Raspberry Pi 5

    This is an INT8-quantized version of OSNet x0.25 optimized for edge deployment on ARM devices using ONNX Runtime, profiled on Raspberry Pi 5. The model was quantized using PT2E static post-training quantization and exported to the .onnx format via the shared PT2E-to-ONNX pipeline for efficient inference on ARM Cortex-A processors.

    image-classification1.1 MBONNX

    Wide ResNet-50 INT8

    Raspberry Pi 5

    An INT8-quantized Wide ResNet-50 image classification model using a widened residual convolutional network for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification72.7 MBExecuTorch

    DeiT-Tiny INT8

    vivo X300SME2

    This is an INT8-quantized version of facebook/deit-tiny-patch16-224 optimized for edge deployment on ARM devices using LiteRT with the XNNPACK backend and KleidiAI acceleration. The model was quantized using post-training static INT8 quantization (PT2E), calibrated on 100 ImageNet-1k images, and exported to the .tflite format for efficient inference on ARM Cortex-A processors (premium smartphone).

    image-classification7.2 MBLiteRT

    SqueezeNet 1.0 INT8

    Raspberry Pi 5

    An INT8-quantized SqueezeNet 1.0 compact convolutional image classification model using fire modules for visual feature extraction. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification1.4 MBExecuTorch

    MobileNetV3-Small INT8

    Alif DK-E8

    An INT8-quantized MobileNetV3-Small image classification model using a lightweight convolutional architecture for mobile and edge inference. Exported as an ExecuTorch program using the Ethos-U delegate for inference on Arm Ethos-U NPUs.

    image-classification2.9 MBExecuTorch

    Swin-Tiny INT8

    vivo X300SME2

    An INT8-quantized hierarchical vision transformer for ImageNet-style image classification, using shifted-window self-attention to capture local and multi-scale visual features efficiently. It accepts 224×224 RGB inputs and produces logits for 1,000 classes, with reduced model size and memory needs compared with the FP32 version.

    image-classification31.7 MBLiteRT

    EfficientViT INT8

    Raspberry Pi 5

    An INT8-quantized EfficientViT image classification model for visual recognition. Exported as an ExecuTorch program using the XNNPACK backend for inference on Arm CPUs.

    image-classification19 MBExecuTorch

    GoogLeNet INT8

    AWS Graviton G4

    INT8-quantized image classification models based on an Inception-style convolutional architecture, designed to reduce model size while preserving accuracy close to the original FP32 variant. They accept 224×224 RGB inputs with standard ImageNet preprocessing and output raw logits for 1,000 classes.

    image-classification6.8 MBExecuTorch

    Develop on Arm with your AI coding assistant

    The Arm AI Portal MCP Server makes it faster to find optimized models, integrate code examples and deploy to your target device from your agentic coding assistant.

    User guide

    Run the following command in your terminal:

    codex mcp add arm-ai --url https://mcp.api.devplatform.arm.com/ai-portal