Microsoft Research has released OptiMind, a 20-billion-parameter language model built for a single, narrow job: taking a business or operational problem described in ordinary language and converting it into the precise mathematical formulation that optimization software can solve. The model is available as an experimental release, with its training data and benchmarks open-sourced.
The bottleneck in optimization
Industries such as energy, finance, logistics and supply chain management have relied for decades on mathematical optimization to schedule crews, route trucks and allocate capital. The hard part has rarely been the solving. It is the modeling: translating a messy real-world situation into decision variables, constraints and an objective function. That work has traditionally required specialists in operations research and can take days or weeks.
OptiMind is designed to do that translation. According to Microsoft Research, the model identifies the decisions to be made, the constraints that limit them and the goal to be optimized, then produces a structured formulation ready for a solver.
Small by design
At 20 billion parameters, OptiMind is small compared with frontier models, but Microsoft says it matches or outperforms several much larger systems on optimization modeling benchmarks. The size has a practical benefit: the model can run locally, so sensitive commercial data does not need to leave a company’s own systems. Microsoft positions that as an advantage for both speed and confidentiality.
Cleaning the training data
A central part of the work, described in the accompanying research paper, was data quality. The researchers found that many public optimization datasets contain errors, ambiguous problem statements and incomplete or incorrect solutions. Rather than train on them as-is, the team classified optimization problems into categories, identified the mistakes most common in each, and used expert-written hints to clean and correct the data. Microsoft says this produced a stronger model and substantially reduced hallucinated formulations.
The same structure is used at inference time. OptiMind first identifies what class of problem it is looking at, then applies the relevant expert guidance when building the formulation. For harder problems it generates several candidate formulations and compares them to arrive at a more reliable answer, an approach that improves accuracy without making the model larger.
What comes next
Microsoft says it is working on reinforcement learning to further improve the model, on generating expert hints automatically rather than by hand, and on integrating OptiMind with everyday business tools such as Excel. The experimental release is meant to gather feedback from researchers and practitioners.
OptiMind is not a general-purpose chatbot and is unlikely to draw consumer attention. Its significance is in a quieter shift: if models like it work reliably, building an optimization model could become a task of minutes rather than weeks, opening a discipline that has been the preserve of specialists to a much wider set of users.
