Two backends can both support weight edits while exposing different parts of the model to the editor.
Cmpnd uploaded dspy-wsp 0.2.0 to PyPI on November 13. The first package version appeared on November 8. The later version expands an experimental DSPy optimizer that lets a separate language model edit a smaller model’s weights. PyPI release · Project changelog
What the release adds
The Clef backend keeps the model backbone fixed and exposes the output component for edits. The Strands backend folds in the checkpoint’s adapter and lets the writer access both the backbone and the output component. The project describes both as MLX backends. Check the selected backend before using the two runs as evidence for a single editing technique. Implementation notes
Which baseline to use
The maintainers report that their Strands Decider 2B run gained 10.4 percentage points from the weight edit alone, 13.7 points from ReAnchor calibration alone, and 20.0 points with both. Against calibration alone, the combination adds 6.3 points in that reported experiment. That difference answers a practical question for teams that already calibrate outputs. It is not an independent replication. For a local comparison, keep the same held-out examples and report baseline, calibration-only, edit-only, and combined scores.
For an adjacent model-serving change, see MustHave’s llama.cpp migration guide.



