What’s New in Open Edge Platform 2026.2 (September 10, 2026)#

The 2026.2 release expands the overall platform offering with a broader set of sample applications, reusable services, and foundational libraries across the edge AI suites. A major theme is the continued shift toward agentic and multimodal workflows, with new capabilities for conversational interaction, video search and summarization, predictive maintenance, classroom assistance, smart kiosk ordering, alert dispatch, enterprise data intelligence, and more.

The platform also strengthens its edge-native execution model through wider CPU, GPU, and NPU enablement, as well as more modular architectures, improved observability and benchmarking, simplified deployment and configuration, and improved documentation to make the platform easier to evaluate, integrate, and scale.

Several old solutions have also been deprecated to prevent bloat and stale offering. These are: Weld Porosity Detection, Worker Safety Gear Detection, Weld Defect Detection, Geti SDK, Chat Q&A, Interactive Digital Avatar, and VLM OpenVINO serving.

For information on specific components, refer to the following sections:

Metro AI Suite#

Metro AI Suite enables scalable Edge AI solutions for cognitive cities, where optimized AI models, microservices, and agents enable developers, ISVs, and solution providers to support public safety, coordinate responses, manage traffic, maintain critical infrastructure, control the environment, and enhance quality of life in real time.

This release enhances developers’ experience in finding the right optimized models, reduces development cycles for multi-modal AI, builds better connections to VMS systems, enhances scene analytics, and enables autonomous monitoring of critical infrastructure assets. This Metro AI Suite release adds and enhances the following features.

Metro Analytics Catalog — is expanded, so you can find the proper pre-optimized AI model, including advanced VLMs that enable real-time scene understanding (28 use cases in total).

AI pipelines — that fuse standard cameras with 3D LiDAR to eliminate blind spots utilizing DL Streamer and SceneScape.

VMS Plug-ins for Video Management System — delivering security alerts directly where operators work.

Critical infrastructure Predictive Maintenance — receives a sample app with Agentic AI enabling outreach to customers in this newly emerging area.

Continued blueprints efforts — showing the art of the possible on Edge AI HW for newly emerging areas, such as Agent Box, Smart Buildings Digital Twin, and Predictive Maintenance across broad range of infrastructure.

See the release notes for specific Metro AI components:

Manufacturing AI Suite#

This release brings defect detection into actionable quality insights through agentic workflows, VLM-based explainability, and ViPPET-powered benchmarking across stream densities and processor configurations. It also adds Windows support in ViPPET to expand deployment flexibility for manufacturing evaluation workflows.

Within Industrial Edge Insights, the Multimodal, Time-Series, and Vision applications deliver explainable weld-quality analysis, faster and more consistent workload evaluation, and more stable, observable pallet and PCB quality inspection. Together, these capabilities help operators act on defect evidence, compare custom computer vision models across stream counts and hardware configurations, and monitor production workloads more reliably.

As part of the 2026.2 streamlining effort, the legacy Weld Porosity Detection, Worker Safety Gear Detection, and Weld Defect Detection sample apps have been retired in favor of unified multimodal and time-series workflows that provide a more scalable path for manufacturing AI deployments.

See the release notes for specific Manufacturing AI components:

Federal and Aerospace AI Suite#

In 2026.2, the Federal and Aerospace Suite continues refining and hardening of Handheld (soldier system) Blueprint, as well as introduces the new UAV blueprint, extending the use of Intel Core Ultra 3 (Panther Lake) as a UAV companion compute, running AI inference workloads alongside the flight controller. The initial release demonstrates object detection with multiple camera feeds and telemetry data from the controller.

See the release notes for specific Federal and Aerospace AI components:

Robotics AI Suite#

This release expands the Robotics AI Suite with stronger support for autonomous mobile robots, humanoid systems, and advanced control workflows on Intel platforms. It adds new localization and navigation options, a ROS-free MPC reference pipeline, and broader support for Intel® Core™ Ultra-based designs, helping teams build more efficient and deterministic robotics systems.

As part of the suite’s redefinition into a more cohesive and comprehensive robotics platform, the documentation has undergone a major overhaul. The updated material makes the suite easier to navigate and deploy across the full robotics workflow, from perception and localization to planning, control, and benchmarking.

The suite now includes new LiDAR-based SLAM backends and expanded reference pipelines for humanoid and mobile robot development, along with additional support for embodied AI and generalist control. These improvements make the platform easier to evaluate and extend while improving consistency across robotics workloads.

For a detailed listing of all the changes, see the release notes.

Retail AI Suite#

This release brings VLM-powered video search enabling visual and timebound descriptions of suspicious activities for loss prevention, combined with intelligent retail assistants that interpret voice cues, accents, and speaker separation for enhanced customer experience and minimized wait times.

Storewide Loss Prevention - Suspicious Activity Detection (1.2.0) introduces Video Search & Recall, allowing investigators to search recorded activity in natural language and instantly retrieve matching video clips. A unified dashboard brings together live alerts, store zone mapping, and video recall in a single interface. Broader hardware support now enables deployment across CPU, GPU, and NPU configurations. The stack upgrades to Scenescape 2026.2.0 while improving first-run model-download handling.

Smart Kiosk Assistant (2026.2.0) delivers a unified platform with dual React-based UIs supporting operator and customer modes. Agentic Ordering with MCP tool calling handles catalog browsing, cart operations, and upsell recommendations. Queue-aware ordering via YOLO-based person counting and RTSP streaming adapts menu recommendations during peak periods. Enhanced voice interaction includes speaker diarization, improved multi-speaker segmentation, and persistent speaker enrollment. Optional multimodal identity supports Face ID and voiceprint authentication. Per-service inference-device configuration enables flexible GPU/NPU acceleration with independent device settings for embedding and re-ranker models.

See the release notes for specific Retail AI components:

Education AI Suite#

Education AI Suite 2026.2 transforms classroom intelligence from single-modal observation to complete multi-modal understanding — combining what’s spoken, seen, and written in real time. With on-device OCR, always-running AI assistants, and expanded developer enablement, this release demonstrates Intel Core Ultra’s ability to orchestrate concurrent AI workloads at the edge without cloud dependency — heterogeneous compute handles OCR, transcription, video analytics, and conversational AI on-device, all at once, preserving student privacy while demonstrating the power of local AI.

See the release notes for specific Education AI Suite components:

Health and Life Sciences AI Suite#

The 2026.2 release introduces a new device category, Surgical Instruments, expanding the existing portfolio alongside Multi-Modal Patient Monitoring and NICU (Neonatal Intensive Care Unit) Warmer.

The Surgical Instruments (initial release) provides a sample application for polyp detection in endoscopic video to evaluate and benchmark edge hardware performance across Intel GPU and NPU, featuring automated YOLO11n training on the REAL-Colon dataset, support for live Basler industrial cameras or recorded video, and real-time observability for latency and utilization metrics.

See the release notes for specific Health and Life Sciences AI Suite components:

Tools and Libraries#

Deep Learning Streamer#

The 2026.2 release solves key edge integration, accuracy, and scaling bottlenecks across following areas.

Open GstAnalytics Metadata — solves vendor lock-in and tool fragmentation by adopting upstream GStreamer standards for tensors and segmentation, eliminating proprietary schemas and custom glue code.

Pluggable Service Extensibility — solves edge hardware capacity constraints and/or complex C++ element development by enabling pipelines to dynamically offload heavy LLM/VLM models to remote servers via simple configuration.

Native RF-DETR Transformer Detection — solves the edge deployment barrier for state-of-the-art transformers by providing out-of-the-box RF-DETR support across Intel CPUs, iGPUs, and dGPUs with zero pipeline redesign.

Live 3D LiDAR + Camera Fusion — solves offline limitations, sensor drift, and linear memory scaling by combining live LiDAR capture, PTS-anchored heterogeneous batching (video, LiDAR, radar), and unified 2D/3D spatial association with in-pipeline rendering in a single graph.

Agentic Developer Experience — solves the steep media analytics learning curve by equipping AI coding agents to generate complete, ready-to-run DL Streamer pipelines directly from natural-language prompts.

Flexible Deployment for Vision Language Models — One GenAI pipeline can run VLM inference locally on Intel hardware or delegate it to a remote OpenAI-compatible server

Scenescape#

SceneScape transitions into a modular, multi-modal spatial intelligence platform, introducing automated 3D sensor fusion, dynamic physical agent support, and persistent cross-scene analytics across the following areas.

Dynamic Physical Agents & Moving Sensors — Solves rigid deployment topologies and sensor-coupling constraints by introducing a unified MQTT contract with publisher-centric topics, enabling mobile platforms (drones, robots, forklifts, UWB/RTLS) to publish observations directly into scenes.

LiDAR + Camera Sensor Fusion Integration — Solves spatial depth blind spots by integrating dual-branch DL Streamer pipelines (3D PointPillars + 2D detection) with the SceneScape Controller, using techniques for unified real-time 3D object tracking.

Automated LiDAR Sensor Calibration — Solves manual multi-sensor setup friction by adding an automated point-cloud localization endpoint to the Auto Calibration Service.

Persistent Re-Identification — Solves attribute loss during tracking interruptions by preserving and merging UUID-keyed historical attributes upon re-identification, while exposing OpenTelemetry latency metrics (average, peak, P95/P99) for cluster sizing and monitoring.

MLOps Model Download Integration — Solves model acquisition friction by integrating the Open Edge Platform Model Download microservice, enabling direct access to external model hubs (such as Hugging Face) with automated OpenVINO IR conversion.

Guided Agentic Deployment Skill — Solves complex multi-step onboarding by equipping AI coding agents with an automated skill that configures pipelines, scenes, and camera parameters with built-in track verification.

OpenVINO#

OpenVINO keeps extending Gen AI coverage and framework integrations to minimize code changes, adding support for several new models and improving multiple currently supported ones. This also means broader LLM model support and more model compression techniques, such as extension of the EAGLE-3 speculative decoding pipeline to LLMs and VLMs, Lazy weight loading, and FP8 quantization.

OpenVINO 2026.3 made a large performance improvement, introducing support for Intel® Xeon® 6+ processors (formerly codenamed Clearwater Forest), and enabling MoE offloading to disk, to run 30B MoE models devices with no more than 16 GB of memory.

OpenVINO™ Model Server simplifies model deployment, reduces command complexity, and further improves stability, performance, and accuracy for LLMs.

For a detailed listing of all the changes, see the release notes for OpenVINO and OpenVINO Physical AI Framework.

Visual Pipeline and Platform Evaluation Tool#

The 2026.2 release enhances ViPPET capabilities with automated consolidated benchmarking suites, platform headroom density testing, and expanded vertical pipelines and model support across the following areas.

In-App Domain Benchmark Suites — solves the friction of manually running multiple tests separately, capturing screenshots, and comparing reports by introducing preconfigured benchmark packs (Retail, Metro, Manufacturing) that execute test cases sequentially across CPU/GPU/NPU and generate consolidated workload- and suite-level performance reports exportable to CSV and PDF.

Extended Stream Density Benchmarking — solves unmeasured platform headroom by enabling users to lock a baseline stream count on one pipeline while automatically scaling a second pipeline to identify hardware limits without compromising primary SLAs.

New Vertical Predefined Pipelines — introduces ready-to-use advanced tracking and time-series pipelines by introducing People Detection & Tracking (combining YOLO, ReID, and DeepSORT) and experimental Wind Turbine time-series anomaly detection.

New model support — improves model coverage by supporting Ultralytics YOLO11 / YOLO26 detection, pose, and segmentation variants.

Sizing Tool Selection Guidelines — provides clear comparison documentation and aligned workflows between ViPPET and the Edge AI Sizing Tool explaining when to use what.

For a detailed listing of all the changes, see the release notes.

Geti™#

Geti™ 3.1 builds on the same lightweight edge-native architecture of Geti™ 3.0, extending it with several new SOTA models while also improving dataset navigation, inference pipeline monitoring, quantization, and other core features.

This release also extends the documentation and it updates some key dependencies (including PyTorch and OpenVINO™) for better performance, stability, and security.

Anomalib Studio#

This release updates the Kaputt dataset source to the official Hugging Face repository, optimizes the binary classification curve metric with PyTorch, switches FastFlow to native padding, and bumps the torch dependency.

It also fixes several issues, including random tiling not honoring the configured tile width, a tar/zip path traversal security vulnerability, invalid handling of seed=0 in random_split, and lingering type-only references to wandb’s removed RunDisabled class.

For a detailed listing of all the changes, see the release notes.

Physical AI Studio#

This release adds support for new policies (RLDX-1, MolmoAct2, and XR0), a new training setup wizard, support for customer-provided robots as plugins to train and deploy models, and improved local and remote workflows for training on GPU — including support for direct trainer URL, Studio-managed SSH connection or optional provisioning of AWS-based training environments.

It also brings the ability to configure LoRA/DoRA fine-tuning for supported policies, image-augmentation directly from the training UI, as well as optimizations and improvements for ACT, Pi0.5, SmolVLA.

For a detailed listing of all the changes, see the release notes.

See the release notes for specific libraries, microservices, tools, and sample applications:

Libraries:

Microservices:

Tools:

Sample applications:

Edge Microvisor Toolkit#

The 26.06.01 patch release of Edge Microvisor Toolkit delivers targeted fixes and package refreshes, including a security fix for Expat and the removal of deprecated support for tink-worker and the device discovery agent. It expands the toolchain with LLVM/Clang/lld 16 support and updates multiple key components for GPU, NPU, and graphics compatibility. It also tightens security, aligns documentation and workflows, and keeps the platform current with the latest supported Intel stack.

For a detailed listing of all the changes, see the release notes.

Image Composer Tool#

Image Composer Tool 2026.2 expands image creation flexibility with faster overlay-based composition, multi-level template extensions, post-boot root filesystem resize, and new support for WSL-compatible Ubuntu images. The overlay workflow allows users to build on existing raw or qcow2 base images for Ubuntu 24.04 and Debian 13, applying only additional or upgraded packages instead of composing the full image from scratch.

This release also adds broader platform and deployment options, including Debian 13 images with custom initrd support and graphical desktop boot using GDM over X11, as well as selective full-disk encryption for raw images using user-provided passphrases. These updates make image composition faster, more modular, and better suited for varied deployment environments while continuing to support minimal image composition from scratch for POR operating systems.

For a detailed listing of all the changes, see the release notes.