Fine-tuning when the base model provider changes

A change of provider is a good moment to check whether the adaptation is still needed. Newer base models often do with a good instruction what an older model learned from examples. According to Menlo Ventures (2025), Anthropic holds 40% of large companies' spending on language model APIs and OpenAI 27%, so companies really do choose between providers, and it is worth having a plan for a change before the first training run.

What the company really carries over

In a hypothetical application, the model prepares short summaries of purchasing correspondence. The team has worked out rules for the records, corrected examples and a set of cases for assessing quality. After a change of provider, it still needs to tell a supplier's offer from a confirmed order.

These agreements are useful material for further work, if the company can use them in the new environment. The adapted model itself is a separate element. Whether its files can be moved, run at another provider or the adaptation reproduced has to be checked for the specific service and licence. Access to the application does not settle these questions.

So it helps to know where the current instructions, the approved examples and the results of earlier comparisons are kept. When they stay only with the contractor, a change of provider may require part of the knowledge about the expected behaviour to be rebuilt. That is work for the team, and it belongs in the decision on migration.

A new model may need a different scope of work

Fine-tuning is further training of an existing model on examples suited to the task. An adaptation done for the previous version does not tell you what training the next one needs. A new general model may already handle some of the old difficulties well and make mistakes elsewhere.

So first assess it with the current instructions and access to the information it needs. In the purchasing correspondence example, the team checks whether the model keeps the status of agreements and whether it leaves out important reservations. Only repeatable errors that remain after such a trial justify assessing further training.

The documentation of Hugging Face model cards says they can describe intended uses, limitations, training data and evaluation results. Such information helps choose candidates to check. A result for the company's correspondence needs its own trial.

Keep the option of a smooth change

The buyer needs information both on the quality of the answers and on the dependencies of the whole application. It may turn out that the model gives good summaries, but connecting the new service requires changes in document handling or in the review of results. The migration scope should keep this work apart from any training.

In the proposed comparison, employees assess the results of both solutions on the same cases that were not used for adaptation. Next to the loss of important information, they also record corrections the new model no longer needs. That lets you judge the change on current work, and saves you from assuming that the previous fine-tuning has to be rebuilt.

For a conversation about migration, a description of the current task, the reason for changing provider and the materials the company has at hand are useful. On that basis we can propose the scope of the assessment and keep the changes the application needs apart from the work on the model. The terms of working together are on the pricing page.

What it costs

Where to start

In the free review we choose one process and 20 to 100 cases the team has already handled. We compare the simpler change with an adapted model on cases the model has not seen, and within 2 working days you receive a written summary with a price. We adapt models as part of private LLMs, and build the application as part of AI apps.