Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Open models turn selection into engineering¶
Meta and Microsoft have announced commercial access to Llama 2, with support across Azure and Windows. The release expands the range of models organizations can host, adapt, and integrate under their own architectural choices rather than consume only through a closed service.
That optionality is valuable. It also moves more of the responsibility from procurement into engineering.
Access changes the control boundary¶
When an organization calls a managed artificial intelligence (AI) service, the provider operates much of the underlying stack. When the organization hosts or adapts an open model, it can gain greater control over deployment location, latency, customization, and data flow.
It also assumes more responsibility for infrastructure, patching, security, evaluation, monitoring, and model lifecycle. The license may permit use; it does not establish fitness for use.
The decision should therefore begin with the control boundary the application requires. Sensitive data, disconnected operations, predictable latency, or specialized adaptation may justify more local control. A low-risk experiment may not justify the operating burden.
“Open” does not mean interchangeable¶
Models differ in architecture, training, documentation, licensing, safety tuning, hardware requirements, and ecosystem support. The label open can refer to several different levels of access and transparency.
Teams should ask specific questions:
- Are weights available, and under which license conditions?
- What is disclosed about training and evaluation?
- Which tools and formats support deployment?
- Can the organization reproduce the tested configuration?
- Which safety behaviors change after fine-tuning?
- Who supplies security updates and operational guidance?
The answers determine whether optionality is real or merely theoretical.
Evaluation must include the operating model¶
A model comparison should not stop at output quality. Total fit includes:
- task performance on representative data;
- compute, memory, latency, and energy requirements;
- security of weights and serving infrastructure;
- ability to monitor and investigate output;
- staff skills needed to operate and adapt the model;
- migration and replacement effort; and
- constraints imposed by the license and deployment environment.
The National Institute of Standards and Technology (NIST) Artificial Intelligence Risk Management Framework is useful because it keeps technical measurement connected to governance and intended use. A model that looks attractive in a benchmark can be a poor organizational choice if the team cannot operate it reliably.
Adaptation creates a new baseline¶
Fine-tuning or otherwise adapting a model can improve performance for a local task. It also creates a distinct system that the organization must document and evaluate. Upstream safety findings may no longer apply completely. Data quality and provenance become part of the model's behavior. A future upstream update may not merge cleanly with the local adaptation.
Maintain a model record containing the base version, training or tuning data, code and parameters, evaluation results, intended uses, known limitations, and owner. Treat every adapted model as a configuration item with a retirement path.
Optionality requires practiced substitution¶
One reason to support multiple models is to avoid hard dependence on a single provider. That benefit exists only if applications separate business logic, prompts, retrieval, and policy from model-specific behavior—and if the organization has actually tested an alternative.
Run substitution exercises. Move a representative workload to another model, identify the failures, and estimate the migration effort. The result will show where supposedly portable architecture has become coupled to one model's quirks.
Llama 2 expands the design space for enterprise AI. The strategic advantage will not come from adding another name to a model catalog. It will come from building the engineering capability to compare, host, adapt, secure, and replace models deliberately.
Open models give organizations more choices. Mature organizations will make those choices with evidence about both the technology and the system they must build around it.
Sources and research trail¶
- Microsoft, “Microsoft and Meta Expand Their AI Partnership With Llama 2 on Azure and Windows” (July 18, 2023).
- Meta, “Llama 2: Open Foundation and Fine-Tuned Chat Models” (2023).
- Touvron et al., “Llama 2: Open Foundation and Fine-Tuned Chat Models” (2023).
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023).
- National Institute of Standards and Technology, Cybersecurity Supply Chain Risk Management Practices (2022).