OpenDPD 2.2.0 — OpenDPD Studio

Try OpenDPD Studio

Launch Studio in your browser →

No installation is needed for the hosted trial. Bring a CSV or try built-in I/Q examples, train PA/DPD models on shared CUDA compute, inspect live plots and download the weights. Temporary workspaces and their files expire within 24 hours.

• CSV uploads are validated in quarantine before preview: UTF-8, two complex or four real I/Q columns, up to 25 MiB and one million paired samples. Rejected and interrupted partial uploads are deleted. Public code, package and model uploads remain unavailable.
• Training now shows separate completed-epoch and current-batch progress bars. Download the best checkpoint while training, then the selected final model afterward.
• Reconnect to running experiments, restore plots and stop an experiment from the read-only terminal.
• English and CUDA when available are the defaults. Nine interface languages, the Studio logo on Home, and improved phone plot controls are included.
• The website and README put the hosted app first; clicking the Studio screenshot opens the app.

Install locally

python -m pip install "opendpd[gui]==2.2.0"
opendpd gui

The wheel includes the frontend; Node.js is not required. Use opendpd[desktop] for the optional native window. The Python API, legacy CLI, metric profiles and checkpoint state-dict format remain compatible.

Full release notes and known limitations · Main changes reviewed in PR #28

Release acceptance used a remote Mac mini (macOS 26.6.2, Chrome 153) against the real HTTPS service: rejected CSV, validated 65,536-sample synthetic CSV, 100 CUDA epochs, separate epoch/batch bars, live/final model downloads and restored plots after reload all passed. Downloaded weights loaded with weights_only=True; the final SHA-256 matched the API. The final batch replay correction is reviewed in PR #29.