What’s New in Open Edge Platform 2026.1 (June 17, 2026)#
Release 2026.1 extends hardware support across Open Edge Platform suites to make Intel® Core™ Series 3 processor (Wildcat Lake) available to their users. It also introduces new additions to the portfolio, with Federal and Aerospace AI Suite, offering an entirely new use case of a Hand-held device using Multi-modal workflows.
Other suites have also been extended with applications such as: NICU Warmer, Storewide loss prevention: Suspicious Activity Detection and Person-of-Interest Reidentification, Smart Kiosk Assistant, Live Video Captioning RAG, WinVision AI, and Metrics Manager. Check out each section for more details.
Importantly, 2026.1 is the final release of Edge Manageability Framework. The project will now be archived and will no longer receive updates. It remains fully available to the open source community for reference.
For information on specific components, refer to the following sections:
Metro AI Suite#
Metro AI Suite 2026.1 accelerates edge AI development with the new Metro Analytics Catalog, a curated collection of pre-optimized AI models and analytics workflows for Intel hardware. It provides an automated script for model downloads and DLStreamer pipelines, significantly reducing time-to-market for Metro AI solutions.
Video Surveillance as a Service sample extends ONVIF camera analytics with an event driven validation pipeline that uses DL Streamer with visual language models to cross check camera events against live video, detect a broader range of objects and context, and provide consistent, explainable analytics across deployments.
This release also introduces enhanced video intelligence with Live Video Captioning RAG, enabling Retrieval-Augmented Generation chat queries against live RTSP video streams using OpenVINO-accelerated LLM inference.
The new Smart Buildings Digital Twin blueprint combines Intel technologies with Scenescape to create high-fidelity virtual building replicas for operational efficiency and predictive maintenance. Security is strengthened with Automated Trusted Compute featuring one-time platform setup that eliminates manual configuration while providing hardware-based root of trust and cryptographic acceleration for secure AI workloads.
Sensor Fusion for Traffic Management blueprint extends to full functionality supported, covering Camera + Rader and Camera + Lidar fusion for Traffic industry. It can support 4x Camera + 2x Lidar fusion running on PTL platform base Edge box, and 12x Camera + 4x Lidar fusion running on ARL + 1x dGPU platform that cover most of the Traffic Management industry requirements.
Importantly, Deterministic Threat Detection is no longer part of the suite.
See the release notes of specific Metro AI components:
Manufacturing AI Suite#
WinVision AI, the first Windows-native AI inference application for the platform is now available, enabling multiple concurrent AI pipelines on CPU, GPU, and NPU with YAML-driven configuration and RTSP/WebRTC output streaming. It is part of the Industrial Edge Insights – Vision collection.
In Industrial Edge Insights – Multimodal, GPU and NPU acceleration for the DL Streamer Pipeline Server and GPU acceleration for the Time Series Analytics microservice are now available, enabling faster weld defect detection inference on Intel hardware.
In Industrial Edge Insights – Time-Series, batch processing UDFs and a new scikit-learn-based ML model for Weld Defect Detection are now available, enabling windowed batch inference and improved detection performance on Intel hardware.
In Industrial Edge Insights – Vision, support for the Intel® Core™ Series 3 processor (Wildcat Lake) is now available across Pallet Defect Detection, PCB Anomaly Detection, Weld Porosity, and Worker Safety Gear Detection.
See the release notes of specific Manufacturing AI components:
Federal and Aerospace AI Suite#
The Federal and Aerospace AI Suite makes its debut in this release, extending the Open Edge Platform offer by a sample application designed for handheld devices.
The Handheld Multi-Modal app enables power-optimized edge deployments with local LLM inference, speech-to-text, a chat UI, and an AI metrics dashboard on SR-IOV-capable hardware. For best results, it is offered with an Edge-Node Infrastructure Blueprint, enabling features, optimizations, and hardware acceleration capabilities. Both software components are considered preview versions.
See the release notes of specific Federal and Aerospace AI components:
Robotics AI Suite#
The Robotics AI Suite 2026.1 release delivers major monitoring, benchmarking, and setup improvements to the Autonomous Mobile Robot component. Level 2 end-to-end pipeline KPI analysis and Grafana live metrics dashboard integration are now available, enabling detailed per-stage latency, throughput, and drop-rate monitoring for ROS 2 deployments. ISX031 industrial camera support and automated one-command environment setup scripts for both Ubuntu Humble and Jazzy are also now available, simplifying hardware onboarding and environment preparation.
See the release notes of specific Robotics AI components:
Retail AI Suite#
The 2026.1 release of Retail AI Suite addresses reliability, accuracy, and scalability for Dine-In and Take-Away Order Accuracy components. The Storewide Loss Prevention suite now introduces expanded behavioral detections, a declarative rule engine, a two-stage VLM-powered analysis pipeline, streamlined auto-generated configuration management, and enhanced performance benchmarking with selectable device profiles to the Suspicious Activity Detection application.
The 2026.1 release updates the POI Re-identification application with a redesigned search API, a two-stage offline search pipeline, multi-embedding detection indexing, enriched entry/exit frame capture, a track purity filter to reduce false positives, and simplified auto-generated configuration management.
This release introduces a new component – Smart Kiosk Assistant - a voice-first retrieval-augmented kiosk stack for retail, QSR, and other customer-facing deployments.
Suspicious Activity Detection#
The Suspicious Activity Detection application now supports a comprehensive set of behavioral detections — including merchandise concealment, checkout bypass, loitering, repeated zone visits, and restricted zone violations. It also introduces a declarative rule engine, which allows rules to be added or modified without any code changes. Building on this, a two-stage behavioral analysis pipeline combines YOLO pose estimation with Qwen2.5-VL-7B-Instruct VLM inference via OVMS to identify concealment in high-value zones, while a session flag system tracks boolean state across zone visits and external service results so that rules can escalate alert severity dynamically — for example, promoting a checkout bypass to CRITICAL when concealment has already been suspected.
Configuration management has also been significantly streamlined, as all environment variables are now auto generated. DLStreamer pipeline configs are generated per camera from the same source, and tracker and reid configs have been moved into each app’s own configuration directory. Furthermore, a new make export-scene target enables exporting scene configurations from a running Scenescape instance as an importable zip file, and performance benchmarking has been enhanced through integration of a new submodule.
Person of Interest Re-Identification#
The POI Re-identification application introduces a redesigned Search API response format, now including entry/exit similarity scores and frame URLs, and historical search has been upgraded to a two-stage offline pipeline that first queries the enrolled POI index before falling back to the all-detections index for non-enrolled persons. Complementing this, the detection index now stores up to five face embeddings per tracked person — captured at ten-second intervals — for more robust matching, while search results have been enriched with per-track entry and exit frames as well as zone-level frames for dwell records.
Configuration management has also been simplified, as all environment variables are now auto-generated, eliminating the need for manual edition. Additionally, a track purity filter has been introduced to prevent false positives caused by DLStreamer track ID reuse by checking per-POI event counts and filtering suspect tracks. Also, benchmark targets have been migrated to use the performance-tools submodule in place of the previous inline backend benchmarks.
Smart Kiosk Assistant#
The initial release of Smart Kiosk Assistant marks the debut of a voice-enabled conversational AI application designed for retail, QSR, airline, and other customer-facing environments, bringing together speech recognition, retrieval-augmented generation, and text-to-speech into a seamless end-to-end voice interaction flow. Built on a unified AI stack that integrates a kiosk UI, orchestration, speech-to-text, retrieval, and speech synthesis, the application grounds every response in an ingestible local knowledge base — delivering context-aware, business-specific answers that feel intuitive and responsive.
Browser-based voice capture and natural audio playback further enhance the user experience, while built-in visibility into model KPIs, runtime details, and latency metrics provides live insight into system performance. Optimized for local and edge deployment, the application utilizes OpenVINO acceleration on Intel hardware for efficient AI inference. Its Docker Compose packaging alongside flexible configuration make it straightforward to deploy, adapt, and scale a cross enterprise environments — establishing Smart Kiosk Assistant as a strong foundation for intelligent, voice-first digital engagement.
See the release notes of specific Retail AI components:
Education AI Suite#
The 2026.1 release of Education AI Suite introduces a major expansion. The Smart Classroom application now offers the Content Search module, a document upload option, OpenVINO-accelerated text and image retrieval, OCR using PaddleOCR, LLM-based Q&A, and multilingual support including Mandarin/Chinese.
WebRTC WHEP streaming for low-latency live video delivery and support for the Intel® Core™ Series 3 processor (Wildcat Lake) are also now available.
Health and Life Sciences AI Suite#
Building on the initial release of the Health & Life Sciences AI Suite, featuring representative and next generation patient monitoring workloads such as AI ECG, remote photoplethysmography (rPPG), and anonymous 3D visual tracking, the 2026.1 release introduces a new neonatal monitoring application.
Neonatal intensive care units (NICUs) are highly specialized environments where early detection and rapid intervention are critical. To support care teams in identifying and triaging acute conditions sooner, medical device manufacturers are increasingly exploring computer vision and multimodal sensing for continuous, non-invasive monitoring.
This release demonstrates a next-generation NICU scenario, where a smart neonatal warmer (patient bed) is equipped with advanced sensing and AI capabilities to enable simultaneous, real-time monitoring:
Object detection (GPU): Environmental and patient context awareness
Contactless vital signs via rPPG (CPU): Continuous heart rate and perfusion insights
Action recognition (NPU): Detection of movement patterns and clinically relevant events
All workloads run within a single, integrated pipeline on one Intel® Core™ Ultra platform, showcasing how heterogeneous compute (CPU + GPU + NPU) can be orchestrated for scalable, efficient edge deployment. With this AI Suite, developers and OxMs can:
Download and evaluate representative workloads aligned to emerging patient monitoring use cases
Validate platform readiness against BOM, performance, and thermal constraints
Prototype multimodal AI pipelines on a unified Intel architecture
This enables earlier alignment with future product requirements and accelerates time to deployment. In addition to the new neonatal workload, the suite includes updates to the original Multi-Modal Patient Monitoring release:
NPU-aware Docker Compose startup
Automatic device detection and configuration
These enhancements simplify deployment and improve portability across Intel platforms, enabling more flexible, automated, and scalable AI solutions at the edge.
See the release notes of specific Health and Life Sciences AI Suite components:
Tools and Libraries#
The Edge AI Libraries 2026.1 release introduces a new speech-to-text microservice, expands platform support, and delivers significant capability enhancements across all components.
Audio Analyzer is now available as a new self-contained OpenAI-API-compatible speech-to-text microservice, enabling edge-optimized transcription with optional voice sentiment analysis on Intel hardware.
DL Streamer Pipeline Server now supports NPU device inference on an Ubuntu 24 base image.
Deep Learning Streamer#
DL Streamer 2026.1 introduces intelligent automation that helps developers build video analytics pipelines faster and more intuitively. The standout addition is an AI coding agent that translates conversational descriptions into working applications. Describe what you need in plain language and get Python code or GStreamer commands ready to run. The agent also assists with migrating existing DeepStream applications to DL Streamer, easing the transition between frameworks. This makes pipeline development accessible to a broader audience while maintaining the power and flexibility the framework is known for.
The framework continues its evolution toward industry standards by fully adopting GStreamer’s upstream metadata APIs. This standardization improves compatibility with the broader ecosystem and simplifies integration with other GStreamer-based tools and workflows.
Several new processing elements expand the framework’s capabilities. Stream multiplexing with proper batch support makes multi-camera deployments more efficient. A new analytics framework provides structured ways to add custom logic for detected objects. Support extends to 3D sensor data, enabling inference on LiDAR point clouds alongside traditional video processing.
Model support grows with the addition of YOLO classification, PaddleOCR for text recognition, and semantic segmentation models. Importing models from popular repositories like Ultralytics, HuggingFace, and PyTorch Image Models become more streamlined, reducing the friction of bringing new models into production.
Windows platform support reaches new maturity with a proper installer, GPU processing through Direct3D, and a much broader set of working samples. The platform now supports the same core workflows available on Linux, making deployment choices more flexible.
Optimizer tools become more sophisticated with capabilities to pause, resume, and inspect optimization in progress. Visualization features of gvawatermark improve with better text rendering, privacy-preserving blur effects, and more control over what gets displayed. Tuning options like CPU thread affinity and selective metadata retention give developers finer control over performance characteristics.
Network camera integration gets easier with dynamic discovery and configuration of ONVIF cameras, now packaged as an installable Python wheel for simplified deployment. The tracking system receives significant refinements for better accuracy in complex scenarios. Sample applications demonstrate practical patterns for common use cases like custom analytics logic, inference optimization, and selective frame processing for vision-language models. Windows platform gains additional samples for YOLO detection and message broker publishing to Kafka and MQTT.
The underlying components stay current with OpenVINO 2026.1, GStreamer 1.28.2, and support for the latest Intel processor generations. The development infrastructure continues to mature with improved testing capabilities and streamlined setup procedures.
Overall, this release focuses on making video analytics development more approachable through AI assistance, more interoperable through standards of adoption, and more capable through expanded model and platform support. The framework continues to serve both rapid prototyping and production deployment needs across diverse hardware configurations.
For a detailed listing of all the changes, see the Release notes for Deep Learning Streamer 2026.1.
Scenescape#
As a critical component for Physical AI applications that require real-world spatial understanding, Scenescape enables AI systems to perceive, track, and interact with physical environments through advanced multi-camera tracking capabilities.
This release adds robust multi-object, multi-view tracking metrics, letting users measure and compare tracking accuracy on real-world 3D positions using industry-standard MOT methods and their own datasets. It introduces occlusion mitigation that uses body pose to keep identities stable through partial blocking, improving reliability in crowded, cluttered scenes.
Re-identification is upgraded to support modern embedding vectors of any size with cosine similarity and explicit track states, enabling more accurate recognition and trustworthy unique counts across people and objects.
The OEP SDK Manager now includes Scenescape with automated discovery, installation, and dependency resolution alongside guided tutorials that eliminate manual configuration steps.
For a detailed listing of all the changes, see the Release notes for Scenescape 2026.1.
OpenVINO#
OpenVINO 2026.2 expands model support with Gemma 4, Qwen3 series, and LFM2 variants, while delivering performance gains through INT4 KV-cache compression, reduced memory usage, and faster model loading for large multi-model pipelines. OpenVINO Model Server adds tool-calling and streaming transcription to support agentic AI workflows.
OpenVINO Physical AI introduces a production-ready framework that standardizes model deployment on robots and camera systems, reducing integration complexity across edge and production environments.
Read more at docs.openvino.ai.
OpenVINO Physical AI Framework#
The first public release of OpenVINO Physical AI Framework introduces a runtime package designed for deploying robot policies in real-world environments, and it packages a core deployment stack that encompasses camera capture, robot interfaces, exported-policy inference, and a runtime loop that connects these components together.
The release features a unified camera API that supports UVC, RealSense, Basler, and shared-camera transport workflows, as well as robot interfaces for SO-101 and Trossen WidowX AI integrations. It also includes an inference runtime for exported policies with built-in OpenVINO and ONNX backends. Furthermore, the package provides a runtime control loop through PolicyRuntime, SyncExecution, and AsyncExecution, along with hardware-specific extras for camera and robot integrations, making it a comprehensive solution for real-world robot inference deployment.
Read more at openvinotoolkit/physicalai
Visual Pipeline and Platform Evaluation Tool#
ViPPET 2026.1 introduces four new predefined pipelines - Video Summarization using Google’s Gemma 3 4B VLM model, Motion Detection using YOLO object detection on motion ROIs, Instance Segmentation using YOLO26n-seg, and Pose Estimation using YOLO26n-pose.
Users can now upload their own video files, image sets, as well as custom OpenVINO models including those trained on Intel Geti platform, what expands pipeline building flexibility beyond predefined assets.
Enhanced performance monitoring now delivers more comprehensive insights. New pipeline latency reporting measures end-to-end pipeline processing time, helping to validate real-time performance requirements. NPU utilization metrics join existing CPU and GPU monitoring.
For a detailed listing of all the changes, see the Release notes for ViPPET 2026.1.
Geti™ Instant Learn#
This release introduces SAM3/SAM3.1 with a modular pipeline, HuggingFace integration, a dedicated OpenVINO backend, quantized model support, and a canvas mode for label-free multi-category segmentation. Cross-image detection with n-shot encoding, PerDINO, Soft Matcher, and Matcher enhancements expand the model offering. On the application side, users gain text prompting, per-project device selection, auto device resolution, frame skipping, a native Tauri file picker, sample dataset management, license enforcement, and a refreshed application icon.
SAM inference latency was reduced across CPU and GPU, while thumbnail generation now respects application quality settings. A mutex guard prevents race conditions during model reloads, a runtime schema validator was added for the MQTT sink, and the stream reader is properly closed on inference loop exit. Unused Triton backends were pruned to reduce image size, the deprecated model optimization parameter was removed to simplify the API, and models were reorganized into self-contained package directories for better discoverability.
Several stability issues were resolved, including a hang on MobileSAM loading, GPU out-of-memory errors in the SAM decoder, a GPU memory leak when switching models, and an OpenVINO Matcher failure on XPU devices. Additional fixes address incorrect bounding box overlay scaling, missing files in stream output, a missing libglib2.0 dependency, incorrect mask merging in MaskedFeatureExtractor, broken MQTT publishing, Windows installer packaging, a bfloat16 tensor crash in TorchModelHandler, thumbnail aspect ratio preservation, post-deletion navigation to the welcome page, camera name handling during stream editing, and Browse button visibility scoped to the Tauri desktop build.
For a detailed listing of all the changes, see the Release notes for Geti™ Instant Learn 2026.1.
Anomalib Studio#
Anomalib version 2.5.0 introduces four new anomaly detection models — INP-Former, GLASS, AnomalyVFM, and CFM — each targeting a distinct detection scenario, including universal, industrial, zero-shot, and multimodal 3D anomaly detection, respectively. To support these additions, a new Hugging Face optional dependency group was added for AnomalyVFM, while the PerlinAnomalyGenerator received notable improvements in the form of a dual-mask mode for richer anomaly shape variation as well as a retry loop to prevent trivially empty augmentations.
For a detailed listing of all the changes, see the Release notes for Anomalib 2026.1.
Physical AI Studio#
Physical AI Studio 0.1.0 is the first public preview of an open source, self-hosted, browser-based application designed to make the full imitation-learning workflow accessible to robotics practitioners, covering everything from hardware setup through dataset recording, model training, and policy execution. It enables users to collect robot demonstration data, train Vision-Language-Action model policies, and run trained policies on their own robot environments, all organized through a project-based structure that manages robots, cameras, environments, datasets, and models.
The application offers guided setup for robot arms, cameras, and reusable recording environments, along with dataset recording, review, import, and export support, as well as model policy training with progress tracking via logs and deployable model exports. Trained models can be executed in-Studio using PyTorch or OpenVINO runtimes, with additional export support for ONNX and ExecuTorch formats. Deployment is streamlined through a Docker-based setup that runs the backend and UI together, while a native development setup is also available for contributors working on either component.
For a detailed listing of all the changes, see the Release notes for Physical AI Studio.
See the release notes of individual components:
Libraries:
Microservices:
Tools:
Edge Microvisor Toolkit#
The 2026.1 release of Edge Microvisor Toolkit delivers two updates. The 25.06 series (version 25.06.03) updates the Linux kernel to v6.12.84 with Intel-specific patches for the drm/xe and IPU6 camera subsystems, syncs the AzureLinux base to 3.0.20260304, updates GPU and NPU drivers, removes the SPECS-EXTENDED and SPECS-SIGNED directories to reduce repository maintenance burden, and tightens the security posture of production images by removing the add-sudoer.sh post-install script from non-development configurations.
For a detailed change log, see Release Notes for Edge Microvisor Toolkit 25.06.03.
The 26.06-preview series introduce Linux kernel 6.18 with Wildcat Lake platform support, upgrades GStreamer to 1.28.1, Mesa to 25.3.4, Intel OneVPL to 2.16.0, and the NPU driver to firmware 1.30.0, and resolves several CVEs in fluent-bit and otelcol-contrib.
See the Release notes for Edge Microvisor Toolkit 26.06 for more details.
Image Composer Tool#
The 2026.1 release of OS Image Composer significantly broadens the scope of supported platforms and deployment scenarios. ARM64 cross-architecture builds are now possible from an x86_64 host, covering Ubuntu 24, eLxR 12, and AZL3 targets. Three new OS distributions are supported: Ubuntu 26.04 LTS, eLxR Edge 26.04 / eLxR 13, and Debian 13.
A declarative network configuration layer in image templates supports both systemd-networkd and netplan, and the attended ISO installer gains an interactive network configuration step. An unattended ISO installer with policy-based target disk selection is introduced as a first-pass implementation. Full offline/cache support for DEB and RPM package metadata and GPG keys eliminates redundant network requests on rebuilds. Additional improvements cover loop device cleanup, partition creation reliability, chroot environment isolation, and mount rollback on failure.
Edge Manageability Framework#
With release 2026.1, the Edge Manageability Framework has been discontinued and will no longer receive updates or active maintenance, including security fixes. Feature requests, bug reports, and pull requests will not be reviewed or responded to. Existing releases and documentation will remain available for reference only.