> For the complete documentation index, see [llms.txt](https://docs.rekor.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.rekor.ai/scout/agent/agent-release-notes.md). # Agent Release Notes This page provides all recent software release notes for Scout Agent as well as links to downloads for both Windows and Linux. ## Agent v6.0.4 Release - 07/31/2026 **What's new in this release** * **Ubuntu 26.04 LTS** — brand-new target, x86-64 and ARM64, CPU and CUDA 13.3. * **Rocky Linux 10** — first RPM-based Linux target, with native `.rpm` packages. * **Expanded CUDA coverage** — GPU builds now span **CUDA 12.6 through 13.3**, including new CUDA 13.0 builds for Ubuntu 22.04. On Ubuntu 22.04 and 24.04, multiple CUDA versions let you match the CUDA version your installed NVIDIA driver already supports instead of upgrading drivers to upgrade the Agent. * **CPU-only builds for every Ubuntu and Rocky Linux target** — run the Agent with no NVIDIA GPU (and no CUDA installation) on any supported Ubuntu or Rocky Linux system. * **ARM64 (aarch64) across the board** — supported alongside x86-64 on all Ubuntu versions, plus NVIDIA Jetson (JetPack 6). * Linux packages are native (`.deb` for Ubuntu/Jetson, `.rpm` for Rocky Linux 10). ### Recognition & Accuracy #### Multi-camera plate grouping restored: The `plate_groups_span_cameras` option — which merges plate reads of the same vehicle seen by multiple cameras into a single group — had been silently non-functional in the 4.x and 5.x release lines. It now works as designed: * With `plate_groups_span_cameras` enabled, a vehicle seen by two or more cameras produces **one** merged group instead of one group per camera. * Grouping now works correctly with **three or more cameras** (previously only the first and last configured cameras could ever share a group). * Fixed partial plate reads occasionally "leaking" into results as separate groups when camera streams were not time-synchronized, and improved tie-breaking so the best plate read is selected for each group. With the option disabled (the default), grouping remains per-camera and is unchanged. #### Virginia plates: letter O normalized to digit 0 Virginia plates use only the digit `0`, never the letter `O` — but recognition results could alternate between the two, causing mismatches in downstream systems. When the plate region is identified as Virginia (`us-va`), the Agent now replaces any letter `O` in the plate result with the digit `0`. This rule requires post-processing rules to be enabled in `alprd.conf`: ``` enable_post_process_rules = 1 ``` #### Updated recognition models: The US OCR model has been retrained with the new state plate designs introduced between 2022 and 2025 for **Alabama, Tennessee, Mississippi, Kansas, Oklahoma, and Pennsylvania**, substantially improving recognition accuracy on those designs. #### Core recognition fixes: Fixed several correctness and memory-safety issues in the core recognition path, improving both result consistency and long-run stability. ### Performance #### Full TensorRT acceleration on Windows (GPU): On Windows systems with a supported NVIDIA GPU, **all** recognition models — including license-plate OCR and vehicle-signature matching — now run on NVIDIA TensorRT by default, rather than only some of them. This increases recognition throughput and reduces latency on GPU installations. *What to expect after upgrading:* the first time the Agent starts, it builds optimized TensorRT engines for your specific hardware. This is a one-time step that can take a few minutes; the engines are cached, so later restarts are fast. #### TensorRT settings now available in the configuration file: TensorRT behavior, previously controlled only through service environment variables, can now be set directly in `openalpr.conf`: | Parameter | Description | | ------------------------------------------ | ----------------------------------------------------------------------------- | | `tensorrt_all_models` | Run all recognition models on TensorRT | | `tensorrt_ocr` | Run license-plate OCR on TensorRT | | `tensorrt_signature` | Run vehicle-signature matching on TensorRT | | `tensorrt_detector`, `tensorrt_classifier` | Jetson only (these models already default to TensorRT on desktop/server GPUs) | | `tensorrt_engine_cache` | Cache built TensorRT engines (default on) | Existing environment variables continue to work and take precedence over the configuration file. Administrators who prefer the previous Windows behavior can disable the all-TensorRT default with `tensorrt_all_models = 0` after removing the setting from the Agent service environment. #### Fixed GPU-accelerated video decode on some H.264 streams: Fixed a regression where GPU (NVDEC) hardware decode could fail for certain H.264 camera streams — the Agent would connect to the camera but process no video. Hardware decode now initializes reliably. #### Improved GPU memory handling: Default batch sizing and GPU initialization were tuned to use video memory more predictably. In addition, the Agent now checks available GPU memory at startup and automatically lowers `gpu_batch_size` if the configured value would not fit, preventing out-of-memory crashes on GPUs with limited VRAM. The configured value is only ever reduced, never raised. ### Reliability #### Automatic recovery from GPU errors: If the GPU or TensorRT runtime cannot initialize on a given system, the Agent now automatically falls back to CPU/CUDA processing and keeps running, instead of repeatedly restarting. Restart loops are now bounded, so a persistent error can no longer put the service into a crash loop. #### Reliable GPU cold starts (Windows): Fixed an issue where a GPU Windows Agent could fail to start on a "cold" system — the first run, before optimized engines had been built. The Agent now prepares all GPU engines before it begins processing video, so startup completes reliably. #### Stalled-pipeline detection: The Agent's watchdog now detects a stalled *recognition* pipeline — not only stalled video input — and restarts automatically to restore processing. #### Crash fixes: * Fixed Agent crashes (segmentation faults) when processing **multiple RTSP camera streams** on Ubuntu 22.04 and 24.04. * Fixed a **long-standing memory leak** that could slowly grow the Agent's memory use over days of continuous operation. * Fixed crashes on shutdown, including heap corruption during teardown on NVIDIA Jetson (JetPack 6) and a crash in `alpr --version`. * Fixed thread-safety issues in GPU recognition that could cause errors or crashes when processing several camera streams concurrently. * Fixed several memory-safety issues in the core recognition path. #### More resilient uploads: * **Upload retry is now enabled by default** (`use_upload_retry = 1`): if the device loses network connectivity, plate reads are queued locally and uploaded once the connection is restored. Previously this had to be enabled manually. * After a reconnect, the Agent now drains the **entire** backlog of queued uploads, not just the most recent items. * Fixed intermittent upload failures (dropped connections) to upload endpoints under load. ### Configuration Tool (alprdconfig) * Fixed video preview getting stuck in **buffering**. * Fixed corrupted plate-region thumbnails when previewing **multiple streams** — random scene fragments could be shown instead of the plate crop. * Fixed a crash when running the built-in **camera test** repeatedly. * The invalid setting `gstreamer_format = rtsp` is now handled gracefully instead of causing continuous camera reconnects. ### Installer * New **Update Rekor Scout Model packages** menu option — update recognition model packages in place, without reinstalling the Agent. * New **Downgrade Rekor Scout Model packages** menu option — roll model packages back to the previous version. * New menu option to **download** the Scout Debian packages without installing them, for offline or staged deployments. ### SDK * The Windows SDK is now provided in two ONNX-runtime variants for broader application compatibility. * A new runtime-only container is available for embedding OpenALPR without the full Agent. * Fixed the **Recognize** API command, which was not working in recent releases. ### Upgrading * No configuration changes are required to upgrade. * **Behavior change:** upload retry (`use_upload_retry`) is now enabled by default, so plate reads are queued during network outages and uploaded on reconnect. Set `use_upload_retry = 0` in `alprd.conf` to restore the previous behavior. * The Virginia O→0 substitution applies only when post-processing rules are enabled (`enable_post_process_rules = 1` in `alprd.conf`). * On Windows GPU systems, allow a few extra minutes on the **first** start after upgrading while optimized GPU engines are built. This happens once and is cached for subsequent restarts. ## Agent v5.1.15 Release - 02/02/2026 Download here: [Windows](https://deb.openalpr.com/windows-agent/openalpr-agent-latest.exe) [Linux ](https://docs.rekor.ai/scout/agent/installation#ubuntu-linux) * Added support for Ubuntu 24.04 Nobile on CPU-only deployments. * Fixed an issue causing errors in the Beanstalk queue. * Introduced new vehicle models. * Improved recognition and handling of California license plates. * Added support for storing video when using file-based video sources. ## Agent v4.1.14 Release 5/28/2025 Download here: [Windows](http://deb.openalpr.com/windows-agent/openalpr-agent-4.1.14-signed.exe) [Linux](https://docs.rekor.ai/scout/agent/installation#ubuntu-linux) * Added timestamps to all log file entries. * Fixed a Windows C# SDK DLL security vulnerability CVE-2024-21907. * Resolved issue where videos could not be found (UVED). * Improved performance for image and video retrieval. * Enhanced PAVE strip identification accuracy. * Fixed an issue where plates were not displayed when store\_plates\_jpeg\_max\_width was set to -1. * Resolved a segmentation fault in 5.1 when running ALPR with GPU enabled. * Added vehicle\_confidence configuration option. ## Agent v4.1.13 Release - 12/4/2024 Download here: [Windows](http://deb.openalpr.com/windows-agent/openalpr-agent-4.1.13-signed.exe) [Linux](https://docs.rekor.ai/scout/agent/installation#ubuntu-linux) * Fixed various GUI issues: * Finish button not working when adding a camera in the Windows GUI. * store\_plates and store\_video settings not editable. * Display settings not resetting when cancel is selected. * Added continuous license validation for offline licenses. * Updated license request process. * License requests moved from to . * Enabled retry logic for PVR uploads by default. * use\_upload\_retry=1 can be configured explicitly if needed. ## Agent v4.1.12 Release - 6/28/2024 Download here: [Windows](http://deb.openalpr.com/windows-agent/openalpr-agent-4.1.12-signed.exe) [Linux](https://docs.rekor.ai/scout/agent/installation#ubuntu-linux) • Fixed an issue where images were not being stored locally.\ • Fixed exclusion masks not applying correctly. ## Agent v4.1.10 Release - 5/16/2024 Download here: [Windows](http://deb.openalpr.com/windows-agent/openalpr-agent-4.1.10-signed.exe) [Linux](https://docs.rekor.ai/scout/agent/installation#ubuntu-linux) * Added fix for clearing all license plates in pvr when configured. * Added ability to pass the strip\_identifying\_information configuration in pvr record. ## Agent v5.0 Release - 5/13/2024 (based on 4.1.9) * Added CUDA 11.4 Support for focal ARM64/AMD64 ## Agent v4.1.9 Release - 5/10/2024 * Lowered default confidence value. * Fixed issue with clearing license plates. ## Agent v4.1.8 Release - 4/8/2024 * Added more International Models. * Post Processing * Remove plates all 1s. * Added ability to clear license plate and state from pvr when configured. * License plate numbers will appear as XX in pvrs * Supports * Edge 300 - configure vehicle\_tracking\_enabled =0 * Jetpack 4.3, 4.4, and 4.6 * Edge Max * Windows * Linux Ubuntu 18.04, 20.04 ARM * Linux Ubuntu 18.04, 20.04 AMD64 ## Agent v4.1.7 Release - 08/2023 * Added US Model improvements for: * Tennessee and Illinois * Model version 5.0 * Supports * Edge 300 * configure vehicle\_tracking\_enabled =0 * Jetpack 4.3, 4.4, and 4.6 * Edge Max * Windows * Linux Ubuntu 18.04, 20.04 ARM * Linux Ubuntu 18.04, 20.04 AMD64 --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.rekor.ai/scout/agent/agent-release-notes.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.