> ## Documentation Index
> Fetch the complete documentation index at: https://imsdkdocs.qualcomm.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Release Notes

> What's new, fixed, and known issues for each QIM SDK release

export const ReleaseSelector = ({children}) => {
  const [selected, setSelected] = React.useState("GA");
  return <div>
      <div style={{
    display: "flex",
    alignItems: "center",
    gap: "1rem",
    marginBottom: "2rem",
    padding: "1rem 1.25rem",
    borderRadius: "10px",
    border: "1px solid #e2e8f0",
    backgroundColor: "#f8faff"
  }}>
        <label style={{
    fontWeight: 700,
    fontSize: "0.9rem",
    whiteSpace: "nowrap"
  }}>Select Release:</label>
        <select value={selected} onChange={e => setSelected(e.target.value)} style={{
    padding: "8px 14px",
    borderRadius: "8px",
    border: "1px solid #d1d5db",
    fontSize: "0.9rem",
    fontWeight: 600,
    cursor: "pointer",
    backgroundColor: "#fff",
    color: "#111827",
    minWidth: "240px",
    boxShadow: "0 1px 3px rgba(0,0,0,0.08)"
  }}>
          <option value="GA">QIM SDK 2.0 GA — Latest</option>
          <option value="rc3">QIM SDK 2.0 RC3</option>
          <option value="rc2">QIM SDK 2.0 RC2</option>
          <option value="rc1">QIM SDK 2.0 RC1</option>
        </select>
        <span style={{
    padding: "3px 12px",
    borderRadius: "20px",
    fontSize: "0.75rem",
    fontWeight: 700,
    backgroundColor: selected === "GA" ? "#dcfce7" : "#fef9c3",
    color: selected === "GA" ? "#166534" : "#854d0e"
  }}>
          {selected === "GA" ? "Latest" : "Older Release"}
        </span>
      </div>
      {React.Children.map(children, child => child && child.props && child.props["data-release"] === selected ? child : null)}
    </div>;
};

<ReleaseSelector>
  <div data-release="GA">
    <Info>
      **Version:** `1.0.2` **Status:** `GA` **Date:** `22nd June 2026`
    </Info>

    ## 1. Reference Kits / Supported Boards

    | SoC platform                 | Reference kits                                                                                                                                             |
    | ---------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
    | QCS6490                      | Qualcomm Dragonwing™ RB3 Gen 2 Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Core Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Kit |
    | QCS5430                      | Qualcomm Dragonwing™ RB3 Gen 2 Lite Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Lite Core Development Kit                                 |
    | Qualcomm Dragonwing™ IQ-9075 | Qualcomm Dragonwing™ IQ-9075 Beta Evaluation Kit (EVK) <br /> Qualcomm Dragonwing™ IQ-9075 Evaluation Kit (EVK)                                            |
    | Qualcomm Dragonwing™ IQ-8275 | Qualcomm Dragonwing™ IQ-8275 Beta Evaluation Kit (EVK)                                                                                                     |
    | Qualcomm Dragonwing™ IQ-615  | Qualcomm Dragonwing™ IQ-615 Beta Evaluation Kit (EVK)                                                                                                      |

    ## 2. Release Overview

    ### 2.1 About this Release

    This is the first major (1.0) release of the Qualcomm IM SDK GStreamer Plugins. It introduces a significant set of new end-to-end sample applications (including Vision-Language-Model inference), a major camera-source plugin update, expanded color-format support, Python bindings, and a large number of stability and build improvements.

    ### 2.2 Release Version

    * **Version:** `1.0.2`
    * **Release Date:** `22nd June 2026`
    * **Release Type:** `Major`

    ### 2.3 Software Dependencies

    * **GStreamer:** `v1.28.2 (Minimum required v1.16)`
    * **External SDKs:**
      * **TensorFlow Lite:** `v2.16`
      * **Qualcomm AI Runtime:** `v2.47`
    * **Camera Service:** `v1.0.2`

    ## 3. What’s New

    ### 3.1 Key Highlights

    * **Vision-Language-Model (VLM) application** powered by the GENIE SDK and the
      Qwen 3B model for natural-language scene and ROI descriptions, with
      multi-stream, IPC, and external WebRTC signalling support.
    * **Seven new AI/vision sample applications** — IP Camera, PPE Compliance,
      Restricted Zone, Gesture Recognition, Object Detection, Video Wall, and Video
      Segmentation.
    * **Camera source plugin generalized** from `qtiqmmfsrc` to `qticamsrc`, with a
      backward-compatible alias (see [Breaking Changes](#4-breaking-changes)).
    * **Expanded video encoder capabilities** — Content-Adaptive Coding,
      Hierarchical B-frames, encoding-mode presets, and bitrate-boost control.
    * **Broader OpenCV color-format support** — I420, YV12, UYVY, YVYU, YUY2, and
      background fill.

    ### 3.2 Enhancements / Improvements

    * **Camera Source (qticamsrc)**
      * **Frame index propagation:** The original camera frame sequence number is now
        propagated to `GST_BUFFER_OFFSET` for downstream debugging.
      * **Runtime capability queries:** Capability queries now use the recorder's
        runtime `FeatureCapabilityMap` instead of a compile-time flag.

    * **Video Transform (qtivtransform)**
      * **Per-frame timestamps:** Frame index and timestamp information is now available
        for every frame to aid debugging.
      * **Video frame origin metadata:** New origin metadata carrying source width,
        height, and any applied crop.

    * **ML Pipeline**
      * **Multi-dimensional audio tensor support:** The audio converter (`mlaconverter`)
        now computes output tensor shape from the selected feature and preprocessing
        parameters, enabling negotiation with multi-dimensional models (e.g., LMFE
        models expecting `[1,1,96,64]`).
      * **Video metadata in mlmetaextractor:** Added `propose_allocation` to advertise
        `GstVideoMeta` support, enabling upstream elements to avoid buffer copies and
        send DMA-buf-backed buffers.

    * **OpenCV Engine**
      * **Expanded color-format support:** Added conversions for I420, YV12, UYVY, YVYU,
        and YUY2 across RGB(A)/BGR(A)/GRAY/NV12/NV21, plus intermediary color conversion
        for packed source formats.
      * **Background fill:** New background-fill capability for blit pipelines and for
        padding when aspect ratios differ.

    * **Platform & Packaging**
      * **Talos platform enablement:** Classification, pose-detection, and face-detection
        sample apps are now enabled on Talos, with automatic SoC-ID detection and
        Hexagon delegate selection (DSP delegate for v66, HTP delegate otherwise).
      * **Upstream/`libcamerasrc` support:** Sample apps fall back to
        `libcamerasrc ! qtivtransform` with NV12 on non-CamX platforms while retaining
        `qticamsrc` with `NV12_Q08C` as the default path. CamX presence is now detected
        by searching for the camera socket, fixing detection inside Docker.

    ### 3.3 New Tools / Utilities

    * **Vision-Language Model (VLM):** Uses the GENIE SDK with the Qwen 3B VLM model
      to produce natural-language descriptions of ROIs and full frames. Supports a
      **standard** use case (periodic or scene-change-triggered descriptions) and a
      **restricted-zone** use case (ROI description only when a person enters a
      predefined area). Supports multiple streams, IPC operation, external WebRTC
      signalling, and builds on Debian and QLI 2.0.
    * **IP Camera:** Single-stage YOLOv8 TFLite object-detection pipeline accepting
      camera, network-stream, or local-file input; streams annotated video over RTSP
      or WebRTC with JSON metadata.
    * **PPE Compliance:** Two-stage ML pipeline (Foot Track + Gear Guard) for
      real-time personal-protective-equipment compliance monitoring.
    * **Restricted Zone:** Real-time analytics pipeline using foot detection plus a
      custom element to determine whether a person has entered a predefined area.
    * **Video Wall:** Single-stage ML object detection across up to 31 input streams
      (USB/ISP/RTSP cameras and MP4), composed into an 8x4 grid with overlays.
    * **Gesture Recognition:** New gesture-recognition demonstration application.
    * **Object Detection & AI pipeline examples:** New YOLO detection examples and a
      collection of C and Python AI pipeline examples (classification, pose
      detection, face detection, detection display).
    * **Camera-via-RTSP example:** Transforms camera streams into an RTSP server
      stream.
    * **RGBA watermark example:** Reworked to use the `qtivoverlay` pipeline,
      supporting one or more watermark images and destination sizes different from
      the image resolution.

    ## 4. Breaking Changes

    * **Camera plugin rename (`qtiqmmfsrc` → `qticamsrc`):** The camera source plugin
      has been renamed to reflect a broader camera-service abstraction. `qticamsrc`
      is now the primary element. Source/header files, library, and package names
      have been updated accordingly.
    * **`mlaconverter` tensor types:** Only `FLOAT32` is now advertised on the source
      caps, as it is the only supported tensor type. Pipelines that relied on other
      advertised tensor types will no longer negotiate.

    ### 4.1 Migration Guidance

    * **Camera plugin:** No immediate action is required. The legacy `qtiqmmfsrc`
      element name is retained as a backward-compatible alias (registered with
      `GST_RANK_NONE`), so existing pipelines continue to function. New pipelines
      should use `qticamsrc`.
    * **`mlaconverter`:** Ensure downstream elements and models consume `FLOAT32`
      tensors. Remove any caps filters relying on previously advertised tensor types.

    ## 5. Bug Fixes

    * **Camera & Capture**
      * Fixed a memory leak in the camera source caused by allocating a context
        containing `std::unique_ptr` members with `g_slice_new0`.
      * Camera pads are now activated during the `READY→PAUSED` transition so that the
        `STREAM_START` event is sent before `CAPS` events, eliminating "Got data flow
        before stream-start event" warnings during `PLAYING→NULL→PLAYING` cycles.
      * Fixed an infinite snapshot loop in the camera prebuffered-data app where a GLib
        timeout callback repeatedly triggered captures every 1 ms.
      * The camera metadata example now supports dynamic capture with metadata, handles
        very long metadata tags without truncation, and no longer exits when the
        USB/ADB controlling terminal is disconnected (SIGHUP/SIGPIPE ignored).

    * **Video Processing**
      * **ib2c / compute shader:** Fixed pixel packing so that stride padding bytes are
        correctly accounted for, resolving incorrect output strides (e.g., 640x360 RGB
        with stride 2560 previously packed as if stride were 1920).
      * **qtivcomposer:** Fixed z-order handling so sink pads are re-sorted when the
        z-order property is updated.
      * **qtivtransform:** Cached surfaces are now flushed on stop and flush events, in
        case the source element has been switched.
      * **qtiobjtracker:** Fixed an unsigned-integer overflow in the `lapjv` function
        (`rowsol`/`colsol` changed to `int32_t`) that could cause segmentation faults.

    * **ML Pipeline**
      * **mlsnpe:** Avoid calling `IBufferAttributesDelete()` on NULL attributes in
        error paths, preventing undefined behavior/crashes.
      * **gst-plugin-ml:** Removed redundant per-frame input tensor-name assignment to
        ML metadata during execute, keeping buffer metadata immutable across the
        QNN, SNPE, and TFLite backends.

    * **Socket Source (socketsrc)**
      * Marked `socketsrc` as a LIVE source to prevent unwanted `_create` calls in
        lower states.
      * Switched from an FD allocator to a DMA-buf allocator so open-source sinks such
        as `waylandsink` receive DMA-buf-backed memory.
      * Resources are now cleaned up during the `PAUSED→READY` transition, fixing
        `PLAYING→READY→PLAYING` cycling.
      * The parent class `change_state` is now invoked before element-specific cleanup
        on downward transitions, following GStreamer's recommended pattern.

    ## 6. Known Issues / Limitations

    * **`mlaconverter` output:** Only `FLOAT32` tensor output is supported (see
      [Breaking Changes](#4-breaking-changes)).
    * **Redis Service** is disabled by default due to power KPI limitations.
    * **IMSDK and Camera service** will work only in permissive selinux builds.
    * The **`gst-ai-superresolution`** sample application may fail to render correctly
      when using file source.
    * Detection accuracy may degrade in the **`gst-ai-face-detection`** sample application
      when faces are far from the camera.
    * The **`gst-ai-pose-detection`** sample application detects the pose of only a single person,
      even when multiple individuals are present in the frame.
    * A red-colored line appears at the bottom of the display when running the **`gst-ai-event-encoder`**
      sample application.
    * A segmentation fault is observed on `Qualcomm Dragonwing™ IQ-9075` and `Qualcomm Dragonwing™ IQ-8275`
      when running the gst-ai-multistream-batch-inference sample application with a batch-8 stream
      using two batch-4 models.

    ## 7. Performance Notes

    * **DMA-buf-backed buffers:** `mlmetaextractor` now advertises `GstVideoMeta`
      support via `propose_allocation`, enabling upstream elements to avoid buffer
      copies and pass DMA-buf-backed buffers, reducing memory-copy overhead.
    * **Pose detection framerate (v66):** Pose detection is capped to 15 FPS via
      `videorate` on v66 Hexagon SoCs to ensure consistent throughput.

    ## 8. Documentation & References

    * [QIM SDK Overview](qimsdk-overview/sdkoverview)
    * [Installation](installation)
    * [Quickstart Guide](quickstart)
    * [Plugin Reference](plugin-reference/introduction)
    * [Sample Applications](sample-application/ai-sample-applications)

    ## 9. Notes & Recommendations

    * **Adopt `qticamsrc` for new pipelines.** While `qtiqmmfsrc` remains available as
      an alias, new development should target `qticamsrc`.
    * **Non-CamX / upstream platforms.** On platforms without CamX, applications
      automatically fall back to `libcamerasrc ! qtivtransform` (NV12). No
      configuration change is required, but ensure `libcamerasrc` is available.
    * **Model and label paths.** Use the `--model-base-path` parameter to point sample
      apps at your model/label directory; defaults are `$HOME` or `/etc/` depending
      on the application.
  </div>

  <div data-release="rc3">
    <Info>
      **Version:** `0.1.4` **Status:** `Release Candidate 3` **Date:** `24th May 2026`
    </Info>

    ## 1. Reference Kits / Supported Boards

    | SoC platform                 | Reference kits                                                                                                                                             |
    | ---------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
    | QCS6490                      | Qualcomm Dragonwing™ RB3 Gen 2 Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Core Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Kit |
    | QCS5430                      | Qualcomm Dragonwing™ RB3 Gen 2 Lite Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Lite Core Development Kit                                 |
    | Qualcomm Dragonwing™ IQ-9075 | Qualcomm Dragonwing™ IQ-9075 Beta Evaluation Kit (EVK) <br /> Qualcomm Dragonwing™ IQ-9075 Evaluation Kit (EVK)                                            |
    | Qualcomm Dragonwing™ IQ-8275 | Qualcomm Dragonwing™ IQ-8275 Beta Evaluation Kit (EVK)                                                                                                     |
    | Qualcomm Dragonwing™ IQ-615  | Qualcomm Dragonwing™ IQ-615 Beta Evaluation Kit (EVK)                                                                                                      |

    ## 2. Release Overview

    ### 2.1 About this Release

    This release fixes major issues across media, camera, and AI pipelines, enabling parallel processing, improving pipeline stability, enhancing camera and encoding support, and adding better ML and TFLite performance along with new utility availability.

    ### 2.2 Release Version

    * **Version:** `0.1.4`
    * **Release Date:** `24th May 2026`
    * **Release Type:** `Minor`

    ### 2.3 Software Dependencies

    * **GStreamer:** `v1.28.2 (Minimum required v1.16)`
    * **External SDKs:**
      * **TensorFlow Lite:** `v2.16`
      * **Qualcomm AI Runtime:** `v2.43`
    * **Camera Service:** `v0.1.1`

    ## 3. Breaking Changes

    ### 3.1 Migration Guidance

    The following plugins are deprecated. Use the corresponding replacement plugins.

    | Deprecated plugins                                                                          | Replacement plugins                                     |
    | ------------------------------------------------------------------------------------------- | ------------------------------------------------------- |
    | `qtioverlay`                                                                                | [`qtivoverlay`](plugin-reference/qtivoverlay)           |
    | `qtimlvdetection`<br /> `qtimlvclassification`<br /> `qtimlvsegmentation`<br /> `qtimlpose` | [`qtimlpostprocess`](plugin-reference/qtimlpostprocess) |

    ## 4. Bug Fixes

    The following issues have been addressed and fixed in this release:

    * **Parallel Processing Support:** Limitations preventing parallel encode or decode sessions have been resolved, enabling simultaneous media processing workflows.
    * **Dynamic Resolution Change (DRC):** Failures in handling files with dynamic resolution changes have been fixed. Buffer allocation between the video driver and GPU driver is now handled correctly.
    * **qtiobjtracker Performance:** The non-linear increase in execution time during long runs has been optimized, improving overall tracking performance and stability.
    * **Audio Recording Pipeline:** An issue where audio pipelines executed without capturing actual audio has been corrected.
    * **GPU + Encoder Pipeline Freeze:** Freezing observed when using GPU-based plugins followed by encoders has been resolved.
    * **Smart Encode with Camera:** Smart encoding with camera sources is now supported, and the NV12\_Q08C format compatibility issue with the encoder has been fixed.
    * **qtivcomposer GAP Buffer Handling:** GAP buffer handling issues have been corrected, improving stability in qtivsplit-based pipelines such as ML daisy-chain use cases.
    * **Camera Service Startup (Config3 & Config4):** Failures in camera service initialization after CamX overlay installation have been resolved.
    * **gst-pipeline-app Utility:** The gst-pipeline-app utility is now available and functional.
    * **sHDRv2/v3 Support (Config2):** Support for sHDRv2 and sHDRv3 camera streams has been added in Config2.
    * **Classification Daisy Chain Pipelines:** Previously unsupported classification daisy-chain use cases are now supported.
    * **Video Encode with UBWC Format:** Encoding for resolutions lower than 4K using UBWC (NV12\_Q08C) format now works as expected.
    * **TFLite Quantized Models (QNN HTP Backend):** Unsupported operations have been addressed. Models now execute reliably, with reduced reliance on CPU fallback.
    * **LiteRT Multistream Support:** Multistream use cases are now supported within the LiteRT framework.

    ## 5. Known Issues / Limitations

    The following are the known limitations in the Qualcomm IM SDK release:

    * In config1, LiteRT IM SDK frameworks don't support the GPU delegate.
    * Frame drops are observed in daisychain detection‑pose usecases when the pose detection load (people per frame × model inference time) exceeds the per‑frame time limit (33 ms for 30 fps).
    * Reverse playback is limited to video streams with a GOP length that fits within the buffer limitations of the video driver (`<=26`).
    * For Dragonwing IQ-9075 and Dragonwing IQ-8275:
      * Random corruption is observed at the beginning of playback for low-resolution (480p) NV12 UBWC-compressed streams.
      * Stability issues are observed when running multibatch ML use cases.
    * The waylandsink plugin doesn't support displaying windows at custom screen coordinates.
    * When using the qtimlpostprocess plugin for postprocessing, a display latency occurs after approximately 30 minutes of running ML use cases on a camera stream.
    * The Qualcomm IM SDK in a docker container isn't supported in this release.
    * The encoder supports only framerates from 10/1 to 120/1 in the order of 10/1. Odd framerates from containers, such as 30000/1001, aren't supported.
    * Across the SNPE, QNN, and LiteRT frameworks, Super Resolution models produce blurred output.
    * Across the SNPE, QNN, and LiteRT frameworks, segmentation models don't function correctly.
    * On Dragonwing IQ-615, the encoder plugin supports only resolutions that are multiples of 16.

    ## 6. Documentation & References

    * [QIM SDK Overview](qimsdk-overview/sdkoverview)
    * [Installation](installation)
    * [Quickstart Guide](quickstart)
    * [Plugin Reference](plugin-reference/introduction)
    * [Sample Applications](sample-application/ai-sample-applications)
  </div>

  <div data-release="rc2">
    <Info>
      **Version:** `0.1.0` **Status:** `Release Candidate 2` **Date:** `26th March 2026`
    </Info>

    ## 1. Reference Kits / Supported Boards

    | SoC platform                 | Reference kits                                                                                                                                             |
    | ---------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
    | QCS6490                      | Qualcomm Dragonwing™ RB3 Gen 2 Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Core Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Kit |
    | QCS5430                      | Qualcomm Dragonwing™ RB3 Gen 2 Lite Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Lite Core Development Kit                                 |
    | Qualcomm Dragonwing™ IQ-9075 | Qualcomm Dragonwing™ IQ-9075 Beta Evaluation Kit (EVK) <br /> Qualcomm Dragonwing™ IQ-9075 Evaluation Kit (EVK)                                            |
    | Qualcomm Dragonwing™ IQ-8275 | Qualcomm Dragonwing™ IQ-8275 Beta Evaluation Kit (EVK)                                                                                                     |

    ## 2. Release Overview

    ### 2.1 About this Release

    QIM SDK 2.0 RC2 is a release candidate focused on expanding AI inference runtime options and broadening platform support. Key focus areas: QNN inference runtime, additional object detection model families, and Ubuntu Server availability.

    ### 2.2 Release Version

    * **Version:** `0.1.0`
    * **Release Date:** `26th March 2026`
    * **Release Type:** `Minor`

    ### 2.3 Software Dependencies

    * **GStreamer:** `v1.28.2 (Minimum required v1.16)`
    * **External SDKs:**
      * **TensorFlow Lite:** `v2.20`
      * **Qualcomm AI Runtime:** `v2.43`
    * **Camera Service:** `v0.1.0`

    ## 3. Breaking Changes

    ### 3.1 Migration Guidance

    The following plugins are deprecated. Use the corresponding replacement plugins.

    | Deprecated plugins                                                                          | Replacement plugins                                     |
    | ------------------------------------------------------------------------------------------- | ------------------------------------------------------- |
    | `qtioverlay`                                                                                | [`qtivoverlay`](plugin-reference/qtivoverlay)           |
    | `qtimlvdetection`<br /> `qtimlvclassification`<br /> `qtimlvsegmentation`<br /> `qtimlpose` | [`qtimlpostprocess`](plugin-reference/qtimlpostprocess) |

    ## 4. Known Issues / Limitations

    The following are the known limitations in the Qualcomm IM SDK release:

    * Parallel decode or encode sessions aren't supported.
    * In config1, LiteRT IM SDK frameworks don't support the GPU delegate.
    * Frame drops are observed in daisychain detection‑pose usecases when the pose detection load (people per frame × model inference time) exceeds the per‑frame time limit (33 ms for 30 fps).
    * Reverse playback is limited to video streams with a GOP length that fits within the buffer limitations of the video driver (`<= 26`).
    * For Dragonwing IQ-9075 and Dragonwing IQ-8275:
      * Random corruption is observed at the beginning of playback for low-resolution (480p) NV12 UBWC-compressed streams.
      * Stability issues are observed when running multibatch ML use cases.
    * The waylandsink plugin doesn't support displaying windows at custom screen coordinates.
    * When using the qtimlpostprocess plugin for postprocessing, a display latency occurs after approximately 30 minutes of running ML use cases on a camera stream.
    * Use cases with files having dynamic resolution change (DRC) may fail. This issue occurs because the video driver allocates buffers that are only partially accessible to the GPU driver.
    * qtiobjtracker exhibits non‑linear execution‑time growth during extended runs, which can act as a performance bottleneck.
    * Classification Daisychain pipeline use cases aren't supported.
    * The Qualcomm IM SDK in a docker container isn't supported in this release.
    * The audio record pipeline runs, but no audio is present in the captured file.
    * Video encode doesn't work for resolutions lower than 4K with the UBWC (NV12\_Q08C) format.
    * Pipelines that use GPU-based plugins followed by the encoder freeze.
    * The encoder supports only framerates from 10/1 to 120/1 in the order of 10/1. Odd framerates from containers, such as 30000/1001, aren't supported.
    * For some models, certain operations aren't supported in the QNN HTP backend. When a CPU fallback is available, the model executes but with degraded performance. If no fallback exists, model execution fails.
    * Smart encode with a camera source isn't supported because, in camera multistream, at least one stream must be in NV12\_Q08C format, and there is a known issue with the encoder for this format.
    * A known issue exists in qtivcomposer related to GAP buffer handling. This affects use cases where qtivsplit is used (for example, daisy-chain ML), as qtivsplit can generate GAP buffers.
    * Across the SNPE, QNN, and LiteRT frameworks, Super Resolution models produce blurred output.
    * Across the SNPE, QNN, and LiteRT frameworks, segmentation models don't function correctly.
    * In the LiteRT framework, multistream use cases aren't supported.
    * In the LiteRT framework, quantized models run at low FPS.
    * In Config3 and Config4, the camera service fails to start after CAMX overlay installation.
    * The gst-pipeline-app utility isn't available.
    * In Config2, sHDRv2 and sHDRv3 camera streams aren't supported.
    * On Dragonwing IQ-615, the encoder plugin supports only resolutions that are multiples of 16.

    ## 5. Documentation & References

    * [QIM SDK Overview](qimsdk-overview/sdkoverview)
    * [Installation](installation)
    * [Quickstart Guide](quickstart)
    * [Plugin Reference](plugin-reference/introduction)
    * [Sample Applications](sample-application/ai-sample-applications)
  </div>

  <div data-release="rc1">
    <Info>
      **Version:** `NA` **Status:** `Release Candidate 1` **Date:** `25th Feb 2026`
    </Info>

    ## 1. Reference Kits / Supported Boards

    | SoC platform                 | Reference kits                                                                                                                                             |
    | ---------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
    | QCS6490                      | Qualcomm Dragonwing™ RB3 Gen 2 Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Core Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Kit |
    | QCS5430                      | Qualcomm Dragonwing™ RB3 Gen 2 Lite Vision Development Kit <br /> Qualcomm Dragonwing™ RB3 Gen 2 Lite Core Development Kit                                 |
    | Qualcomm Dragonwing™ IQ-9075 | Qualcomm Dragonwing™ IQ-9075 Beta Evaluation Kit (EVK) <br /> Qualcomm Dragonwing™ IQ-9075 Evaluation Kit (EVK)                                            |
    | Qualcomm Dragonwing™ IQ-8275 | Qualcomm Dragonwing™ IQ-8275 Beta Evaluation Kit (EVK)                                                                                                     |

    ## 2. Release Overview

    ### 2.1 About this Release

    QIM SDK 2.0 RC1 is the initial 2.0 release candidate, introducing the new plugin architecture and expanded AI inference runtime support.

    ### 2.2 Release Version

    * **Version:** `NA`
    * **Release Date:** `25th Feb 2026`
    * **Release Type:** `Minor`

    ### 2.3 Software Dependencies

    * **GStreamer:** `v1.28.2 (Minimum required v1.16)`
    * **External SDKs:**
      * **Qualcomm AI Runtime:** `v2.41`

    ## 3. Breaking Changes

    ### 3.1 Migration Guidance

    The following plugins are deprecated. Use the corresponding replacement plugins.

    | Deprecated plugins                                                                          | Replacement plugins                                     |
    | ------------------------------------------------------------------------------------------- | ------------------------------------------------------- |
    | `qtioverlay`                                                                                | [`qtivoverlay`](plugin-reference/qtivoverlay)           |
    | `qtimlvdetection`<br /> `qtimlvclassification`<br /> `qtimlvsegmentation`<br /> `qtimlpose` | [`qtimlpostprocess`](plugin-reference/qtimlpostprocess) |

    ## 4. Known Issues / Limitations

    The following are the known limitations in the Qualcomm IM SDK release:

    * AI/ML parallel inference for 32 streams varies between 20–30 fps depending on the model and input stream, which is less than the expected 30 fps.
    * Frame drops are observed with the Qualcomm® Neural Network plugin when running on GPU delegate.
    * Frame drops are observed in daisychain detection‑pose usecases when the pose detection load (people per frame × model inference time) exceeds the per‑frame time limit (33 ms for 30 fps).
    * Reverse playback is limited to video streams with a GOP length that fits within the buffer limitations of the video driver (`<= 26`).
    * For Dragonwing IQ-9075 and Dragonwing IQ-8275:
      * Random corruption is observed at the beginning of playback for low-resolution (480p) NV12 UBWC-compressed streams.
      * Stability issues are observed when running multibatch ML use cases.
    * The waylandsink plugin doesn't support displaying windows at custom screen coordinates.
    * When using the qtimlpostprocess plugin for postprocessing, a display latency occurs after approximately 30 minutes of running ML use cases on a camera stream.
    * Use cases with files having dynamic resolution change (DRC) may fail. This issue occurs because the video driver allocates buffers that are only partially accessible to the GPU driver.
    * qtiobjtracker exhibits non‑linear execution‑time growth during extended runs, which can act as a performance bottleneck.
    * ROI isn't being generated in daisychain detection‑classification use cases when using a USB camera source with the YUY2 format.
    * The qtiqmmfsrc and qtimltflite plugins aren't enabled in this release.
    * The Qualcomm IM SDK in a docker container isn't supported in this release.

    ## 5. Documentation & References

    * [QIM SDK Overview](qimsdk-overview/sdkoverview)
    * [Installation](installation)
    * [Quickstart Guide](quickstart)
    * [Plugin Reference](plugin-reference/introduction)
    * [Sample Applications](sample-application/ai-sample-applications)
  </div>
</ReleaseSelector>