Open weight models have quietly become the enterprise default for a specific class of workload.

The debate about whether open weight models would find a durable place in the enterprise stack is largely resolved. They have. The interesting question is which workloads they own and which they do not.

In interviews with sixteen enterprise machine learning leads over the past quarter, the same rough split emerged. Complex reasoning, agentic planning and long context synthesis remain the territory of the frontier closed models. Classification, extraction, routing, moderation, first pass summarisation and the retrieval side of retrieval augmented generation are increasingly served by fine tuned open weight models running on the enterprise's own infrastructure.

The economic logic is straightforward. On a high volume workload, the marginal token cost of a hosted frontier model dominates the total cost of an application. On a low volume, high value workload, model quality dominates. Most enterprise portfolios contain both kinds of workload, and the tooling has matured enough to route between them.

The vendor implication is that the closed model providers are converging on a business that looks less like the model layer and more like the reasoning layer. That is a smaller market than they were originally underwritten against. It may be a more defensible one.