FreeLLMCmpnd launches dspy-wsp, a gradient‑free weight‑space optimizer for DSPy decision models
Cmpnd released dspy-wsp, a prototype DSPy optimizer that edits model weights directly using GPT‑6‑Astra, improving decision‑model accuracy without gradients.
The tool implements weight‑space programming, letting a frontier LLM such as GPT‑6‑Astra read and edit another model’s weights without gradients or training data. It is packaged as the WeightSpace optimizer, works with any DSPy program and metric, requires DSPy 3.5, and is available under an MIT license via pip.
In tests on five open decision models ranging from 0.6 B to 9 B parameters, each run used about $5–10 of GPT‑6‑Astra calls and 25 forward passes. All models improved; d1‑omni‑600M increased from 50.9 % to 69.1 % when WeightSpace was combined with ReAnchor, a gain of 18.3 points. The 9 B Clef‑Flash model saw a smaller rise of 1.9 points with WeightSpace alone and 6.5 points with ReAnchor.
The experiments ran on a single Mac, with 2 B models using roughly 6 GB of memory and the 600 M model needing at least 4 GB. Each optimization took between 78 and 167 minutes. Linux users can run the optimizer on PyTorch with an NVIDIA GPU.
The authors describe the tool as "very experimental" and note it is slower on larger models, while editing text‑generating LLMs remains costly.
What it means for you
For developers, dspy-wsp provides a cheap, gradient‑free method to boost decision‑model accuracy while keeping the model on local hardware. For users who prefer free or self‑hosted models, the release shows that weight‑space edits can be applied on modest machines, but the technique is still early and may not work well for larger models.
Sources
- Cmpnd Releases dspy-wsp: Weight-Space Programming Comes to DSPy, MarkTechPost, 14 Nov 2026
- Weight-Space Programming with DSPy — Cmpnd, cmpnd.ai (official), 13 Nov 2026
Written automatically from the sources above, then checked by our code: every number and name must appear in them, and no sentence may copy a source. Spotted a mistake? Report it.
Which AI can you use for free today?
Every free AI model, ranked by blind human votes, with the memory and speed it needs on your own GPU.