Home›Services›Custom LLMs

Service

Custom LLM fine-tuning

Fine-tuning is the most over-sold service in this industry. It is genuinely the right answer perhaps one time in five, and we would rather be the firm that tells you which time it is than the one that sells it to everybody.

Try these three things first

1. A better prompt and a few real examples. Most quality complaints are solved here, in an afternoon, for nothing.
2. Retrieval. If the model does not know your products, your policies or your case history, the fix is to give it access at the moment of the question — not to bake a snapshot into weights that will be out of date next quarter.
3. A bigger model for the hard ten per cent. Routing the difficult cases to a stronger model is usually cheaper than training your own.

When fine-tuning is actually right

• Volume. Millions of similar calls, where a small tuned model matches a large one and costs a fraction to run.
• Format discipline. Output that must follow a rigid structure every time, where prompting gets you to ninety-five per cent and the last five per cent matters.
• House style. A voice or a document convention that is genuinely yours and cannot be described in a paragraph but can be shown in a thousand examples.
• Isolation. Data that may not leave your premises, so the model must run where the data already is.
• A narrow language or domain. Where general models are weak and your corpus is strong.

How we run it

We start with the evaluation set, not the training data. Until there is a way to measure whether the tuned model is better than what you have, training is an expensive way to generate opinions. The set is built from your real cases and includes the ones your current system gets wrong.

Then data preparation, which is most of the work: cleaning, de-duplication, removing personal data that has no business being in weights, and balancing so the model does not simply learn your most common case. Training itself is usually the shortest step. Afterwards the tuned model, the base model and the frontier model all run against the same set, and you see the three numbers side by side — including the times the tuned model loses.

Self-hosting

We run open models in production ourselves, on our own hardware, for exactly the workloads described above. That means we can tell you the real numbers: what a machine costs, how many concurrent requests it handles, where the latency actually goes, and how it behaves at three in the morning when nobody is watching. For high-volume work the per-request cost is effectively zero, which changes what is worth automating at all.