OptiMind is a specialized Microsoft Research language model that translates natural language optimization problems into solver-ready mathematical formulations, generating both the formulation and executable GurobiPy code so solvers can run MILPs directly from user descriptions.

Led by Doug Burger, Managing Director of Microsoft Research Core Labs, OptiMind converts plain-language descriptions into formal optimization formulations to facilitate exploration of solutions with powerful optimization solvers.

The project envisions long-term applications to larger systems—cities, infrastructure, and local economies—with potential contributions to sustainability by reducing emissions.

The initiative emphasizes openness and accessibility through open-source exploration on Hugging Face.

In benchmarking, OptiMind shows about a 20.7% gain in formulation accuracy over the base model, with test-time scaling techniques narrowing gaps to larger proprietary models under evaluation.

Inference follows a multi-stage pipeline: classify inputs into 53 optimization classes, augment prompts with class-specific hints, generate reasoning traces and a final formulation plus GurobiPy code, with optional self-consistency voting and multi-turn corrections.

The model aims to lower barriers to advanced optimization modeling, enabling faster experimentation, iteration, and learning for researchers and practitioners.

The release sits within a broader EdTech and innovation context, as evidenced by the ETIH Innovation Awards 2026 recognizing measurable impact in education technology.

This effort is part of democratizing optimization through generative AI and agentic solutions, combining LLMs with simulators and existing optimization algorithms.

Primary use cases include supply chain network design, manufacturing and workforce scheduling, logistics and routing with real-world constraints, and financial portfolio optimization.

The base model is openai/gpt-oss-20b, fine-tuned on cleaned optimization datasets and released under the MIT license, with evaluation on IndustryOR and Mamo Complex benchmarks.

Getting started resources include trying OptiMind on Hugging Face, using Microsoft Foundry for experimentation, and consulting the Microsoft Research blog for technical details and results.