Beyond Public Models: The Return of Enterprise Control

Welcome to a new edition of The Board: Distillation Aftershots ^(*).

This online version of the newsletter, written by Constellation's Chief Distiller and Board Advisor Esteban Kolsky, shares curious and interesting insights and data points distilled from enterprise technology to identify what’s notable. If you'd like to get this in your inbox, subscribe here.

The discussion about open-source AI has inherited too much baggage from the software industry. Most executives, technologists, and advisers understand open source through Linux, databases, infrastructure software, and applications. The familiar model starts with publicly available code, allows organizations to modify or fork it, and eventually produces commercial versions that add support, security, stability, and enterprise services. That history is useful, but it also makes the current discussion about AI harder than it needs to be because models introduce another option: open weights.

First, my take.

Open-source software gave enterprises access to code. Open weights give them access to the learned parameters that determine how a model responds. This is a critical distinction because enterprises need control over model behavior, deployment, economics, and proprietary context rather than access to another general-purpose application. The Open Source Initiative defines a model as the architecture, parameters, including weights, and inference code, while its definition of truly open-source AI also requires the information and code needed to derive those parameters. An open-weight model gives the enterprise substantial freedom to customize the model without necessarily exposing the complete training process that created it.

This becomes more relevant as the enterprise AI strategy changes. The first several years of generative AI adoption were dominated by public and frontier models because they were immediately accessible and demonstrated capabilities that enterprises could not economically reproduce themselves. Public models were useful for experimentation and generic activities, but their limitations become more visible as organizations try to move into proprietary processes, differentiated decisions, confidential data, and production-scale economics. McKinsey now argues that organizations with well-structured internal data can lower technology investment by fine-tuning smaller, domain-specific models, with potential advantages in cost, resilience, and compliance. Deloitte similarly finds that open and derivative models are increasingly competitive with frontier models and may be better suited to specialized enterprise workloads.

The performance argument for defaulting to the largest closed model is also getting weaker. Stanford’s 2025 AI Index found that the benchmark gap between the leading closed- and open-weight models had fallen from about 8% to 1.7% in little over a year. Model leadership will continue changing, but the broader conclusion remains: enterprises now have credible alternatives across a growing range of workloads.

This connects to a point I have been making throughout the past year about public models and enterprise value. General-purpose models are optimized to serve a very large population and a very broad range of problems. Enterprise advantage comes from the opposite direction: proprietary context, privileged data, institutional knowledge, workflow ownership, and an understanding of the decisions that matter inside a particular organization. The closer AI moves toward those assets, the less useful the assumption becomes that the largest public model should remain at the center of the architecture.

Open weights create an interesting middle ground. An enterprise can start with a capable model whose parameters are available, fine-tune or adapt it around its domain, deploy it in an environment appropriate to its security and latency requirements, and control more of the resulting economics. Meta, for example, explicitly supports downloading Llama weights, training them on new data, fine-tuning them, and running them on-premises or in other environments without sending enterprise data back to Meta. Google’s Gemma family provides pretrained open-weight models in multiple sizes specifically designed for further fine-tuning, while smaller variants demonstrate how specialized models can reduce infrastructure requirements substantially.

This does not mean every enterprise should train its own foundation model. For most organizations that would recreate the economics they are trying to escape. The more interesting opportunity is to create a portfolio of models. Public frontier models can continue serving problems where breadth and rapidly changing capabilities matter. Smaller proprietary or domain-specific models can serve repeatable workloads where context, confidentiality, latency, economics, and predictable behavior matter more. Open-weight models provide another source from which those enterprise-specific capabilities can be developed.

That portfolio approach is already showing up in the move toward agentic AI. Deloitte reports that 85% of companies expect to customize agents around the needs of their businesses, while McKinsey argues that off-the-shelf agents are unlikely to create strategic advantage in high-value processes because those processes depend on company-specific logic, data flows, and value drivers. Their recommendation is increasingly modular: combine custom and commercial components while avoiding architectures that hardwire the enterprise to one platform.

This is where the old open-source analogy to software becomes limiting: the question is no longer whether someone can inspect the source code or whether a community can improve a product. The useful question is how much control the organization needs over the intelligence embedded in its operating environment.

Open weights can provide control over customization, deployment, data exposure, model behavior, and eventually economics without requiring the enterprise to become a frontier-model laboratory.

Leading enterprises are beginning to understand that their advantage will come from what they place around and inside the models: proprietary data, context, semantics, workflows, governance, institutional experience, and the ability to select and modify different models for different problems.

Open weights can reduce dependency on a single provider, limit lock-in, and protect the privileged operating logic that makes the organization different.

An enterprise using open-weight or customized models takes on more of the model lifecycle itself: evaluation, security testing, fine-tuning, version management, governance, monitoring, and decisions about when to adopt improvements from newer upstream models. Open weights are therefore not a shortcut around infrastructure or operations. Their value increases when the enterprise has a platform capable of managing models as long-lived assets rather than one-off experiments.

For CEOs and boards, the next discussion should include model ownership and control as part of infrastructure strategy. Determine which workloads can remain on public models, where domain-specific models create better economics or performance, which enterprise data and context should be incorporated into customized models, and how much dependency on any single model provider is acceptable. The objective is to avoid model lock-in while preserving the privileged data, context, workflows, and operating decisions that make the enterprise unique. Those assets should remain portable across models and providers even as the underlying technology changes.

The strategic value of open weights is giving enterprises another path to preserve control over proprietary context, operating logic, deployment choices, and provider dependency as the model market continues to change.

Here are some reading resources:

• Open Source Initiative. “The Open Source AI Definition 1.0.” The clearest formal distinction between models, weights, source code, and the information required for truly open-source AI.
• Stanford HAI, “2026 AI Index — Technical Performance.” Tracks the continuing convergence and divergence between closed and open models and shows how quickly model leadership changes.
• McKinsey, “Building the Foundations for Agentic AI at Scale.” Explains why smaller domain-specific models can reduce cost while improving resilience and compliance.
• McKinsey, “Seizing the Agentic AI Advantage.” Makes the case for combining custom-built and commercial capabilities around company-specific processes and data.
• Deloitte, “Enterprise AI Infrastructure Survey: A 2028 Outlook.” Examines the growing role of open and derivative models in specialized enterprise workloads.
• Meta, “Introducing Llama 3.1.” Useful for understanding the practical freedoms available when model weights can be downloaded, fine-tuned, and deployed privately.
• Google, “Introducing Gemma 3.” Shows the growing range of open-weight models intended for domain customization and different infrastructure footprints.

What’s your take? We are fostering a community of executives who want to discuss these issues in depth. This newsletter is but a part of it. We welcome your feedback and look forward to engaging in these conversations.

If you are interested in exploring the full report, discussing the Board’s offering further, or have any additional questions, please contact me at [email protected].

(*) A normal distillation process produces byproducts: primary, simple ones called foreshots, and secondary, more complex and nuanced ones called aftershots. This newsletter highlights remnants from the distillation process, the “cutting room floor” elements, and shares insights to complement the monthly report.