Fine-Tuning Quietly Turned Into Fleet Management

The mental model most organisations still carry is that customising a language model means training a new one: an expensive job, run rarely, producing a large artefact that is deployed and versioned like a release. Parameter-efficient adaptation broke that picture. Instead of updating every weight, these techniques train a small additional set — a fraction of a percent of the model's size — that modifies behaviour when attached to an unchanged base model. The resulting artefact is small enough to store casually and cheap enough to produce in hours on modest hardware. The operational consequences are more interesting than the technical ones. Because adapters are small and share a base, many can be served from one deployment, swapped per request, and even composed. A team that would previously have run one general-purpose model can now run a per-customer, per-domain or per-task adapter without proportional infrastructure cost. That is a genuine capability gain, and it changes what is worth customising: work that could never justify a training run now can. It also creates a governance problem that arrived faster than the practices to handle it. When adaptation was expensive, the number of model variants in production was small enough to hold in your head. When it costs almost nothing, variants proliferate, and organisations quickly reach a state where they cannot answer basic questions: which adapters exist, what data each was trained on, which requests route to which, and whether any of them still reflects a policy that has since changed. Each adapter is a behavioural change to the system with no obvious place in the change-control process, because it does not look like code and does not look like configuration. The teams handling this well treat adapters as first-class deployable artefacts — registered, versioned, tied to a training dataset with recorded provenance, and subject to the same evaluation before promotion as any other change. The teams handling it badly are not doing anything reckless; they are simply producing artefacts faster than their inventory process was designed to track, which is how most infrastructure sprawl has always started.