# 0.12.0

## Release Information

- OS support: Tested with Windows-x86 (10), Windows-arm64 (10), Mac-arm64 (14), Linux-x86 (Ubuntu 22.04, Ubuntu 24.04), Linux-arm64 (Ubuntu 24.04)
- QAIRT SDK: Tested with versions (2.49.0, 2.50.0)
- Model Visualization format support: ONNX, TensorFlow, PyTorch (1.x, 2.x), TFLite, Executorch Program, DLC, EAIX, HTP, and QNN JSON

## Highlights

- DLC Inspector

    - Added support for inspecting the contents of a DLC file when a DLC file is visualized, letting users browse the ONNX or HTP records embedded in a DLC and open them in a window, or download them
- HTP Op Mapping support for Mac

    - Extends the Op Tensor Mapping feature to Mac (arm64) for DLC models compiled with HTP context binary generation
- Undocked workspaces

    - Undocked workspace windows are now reopened automatically the next time the application is launched, alongside the main window’s workspaces
    - Op/Tensor Mapping, Model Diff, subgraph windows, and windows launched from the CLI or Python API are excluded from persistence
    - Undocked workspaces can now be dragged back to the main application window, with a visual drop-zone highlight indicating valid targets
    - A workspace can be redocked by dragging its tab back into the highlighted tab zone of the main window
- Graph side panel revamp

    - Redesigned layout of the graph side panel properties view
    - Input and output section producers and consumers now have a locate action next to each that scrolls the graph to that node without changing the current selection node properties selection
- Quantization tab rework

    - Rebuilt the quantization encodings view with clearer per-tensor formatting and provided search across tensor names and encoding types

## Bug Fixes

- Fixed a memory issue where restoring or undocking a workspace with large model could exhaust memory
- Fixed a memory leak where closing a graph panel or its workspace did not fully release its resources
- Fixed accuracy report sorting when metric columns contain NaN values

## Known Issues

- Diffing larger models in the QAIRT Visualizer Graph Panel (i.e., with 40K nodes or more or larger than 1.5 GB) may result in performance degradation or may fail to load altogether due to memory constraints. Upcoming releases will address these limitations to allow extra-large models to be loaded.
- Models which have individual weights that are larger than 2 GB will fail to load. A fix is coming in an upcoming hotfix.

## Notices

- QAIRT Visualizer content may include China design and development contributions

Last Published: Sep 14, 2026

[Previous Topic
Notices](https://docs.qualcomm.com/bundle/publicresource/80-87189-1/topics/0_11_0.md) [Next Topic
0.12.1](https://docs.qualcomm.com/bundle/publicresource/80-87189-1/topics/0_12_1.md)