Particle.news
Download on the App Store

Microsoft Research Releases OptiMind, a Model for Solver-Ready Optimization From Natural Language

The public-preview release invites community testing of a MILP-focused system that generates solver-compatible code through a structured pipeline.

Overview

  • OptiMind is now available in experimental form on Microsoft Foundry and on Hugging Face for open exploration and integration.
  • The model targets the formulation step by converting plain-language problem descriptions into mixed integer linear programming models.
  • Its multi-stage workflow includes problem classification, hint retrieval, generation in formats such as GurobiPy, and optional self-correction.
  • Microsoft reports internal evaluations on cleaned benchmarks like IndustryOR, Mamo-Complex, and OptMATH with roughly a 10% accuracy gain over the base model and competitive results versus larger systems.
  • Resources for practitioners include a research blog, a technical paper, playground access, and sample code on GitHub, with feedback encouraged during the preview.