Microsoft Research reposted this
Optimization has always sat quietly underneath the most important industrial decisions. Pricing, supply planning, production, logistics. The math has been there for decades, but using it has rarely been simple. In CPG and retail, I have watched teams spend years translating real-world constraints into spreadsheets and hand-built models. It works, but only just. The gap between how people think about the business and how optimization systems need to be specified has always been the hard part. Today we are launching Microsoft Research's OptiMind in Microsoft Foundry Labs to explore a different approach. OptiMind is a 20b parameter model, trained deeply in the structure of optimization problems. You describe the situation in plain language. The model turns that description into a formulation that can actually be solved. Because it is trained for this specific kind of work, it tends to be more reliable than much larger general-purpose systems. And because it can run locally, sensitive operational data stays where it belongs. For industries where constraints are real, this changes the shape of the work. Optimization becomes less about specialist tooling and more about everyday reasoning.