LIVE

Cmpnd Launches dspy-wsp to Introduce Weight-Space Programming to DSPy

🤖 AI Models 2 min read 👁 9

Developer tool startup Cmpnd has released dspy-wsp, a prototype optimizer that enables frontier language models to directly inspect and edit the weights of smaller decision models without traditional gradient training.

Cmpnd Launches dspy-wsp to Introduce Weight-Space Programming to DSPy🤖AI Models

Photo: MarkTechPost

Key points

  • Cmpnd released dspy-wsp under an MIT license to integrate weight-space programming into DSPy.
  • The technique uses a frontier model to directly edit student model weights without gradient training loops.
  • Tests on models ranging from 0.6B to 9B parameters showed variable performance gains, requiring local execution on Mac hardware.

Developer tool startup Cmpnd has introduced dspy-wsp, an open-source optimizer that brings weight-space programming (WSP) directly into the DSPy ecosystem. Released under an MIT license, the package allows a frontier model to read and modify another model's parameters directly, bypassing conventional training loops, gradient descent, and fine-tuning datasets entirely.

According to the release details, the underlying WeightSpace optimizer integrates with any DSPy program and evaluation metric. In testing across open decision models ranging from 0.6 billion to 9 billion parameters, the method required about five to ten dollars in frontier model API calls per run. Experiments were conducted locally on standard Mac hardware, with smaller models like the 600M parameter variant consuming around 4 GB of memory and larger 2B models using approximately 6 GB.

Performance outcomes varied across architectures and configurations. When paired with a technique called ReAnchor, smaller models registered notable accuracy boosts, with one 600M model jumping from 50.9% to 69.1%. However, the approach proved less impactful on larger models, such as a 9 billion parameter variant that saw marginal gains, and current runs remain experimental and computationally slow, taking between 78 and 167 minutes to complete.

What to watch next: Monitor whether developers adopt weight-space patching as a viable alternative to traditional fine-tuning for smaller open-source models.

Source MarkTechPost — marktechpost.com/2026/11/14/cmpnd-dspy-wsp-weight-space-programming/ ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.