Blog: LLM Fine-Tuning Notes | LLM Fine-Tuning Services

What does it cost to host a fine-tuned model?
That turns on the billing unit for serving. Some charge tuned models per token at or near base-model rates; others put tuned models on dedicated GPUs billed by the second, minute or hour whether or not requests arrive. At low traffic, per-token serving is usually far cheaper. At steady high traffic, dedicated capacity can win.
Costs, last revised November 7, 2026
How much data do you need to fine-tune an LLM?
For a narrow task with a clear right answer (classification, extraction, a fixed output format), a few hundred checked examples is often enough to see whether fine-tuning helps. Broader tasks need thousands. Whatever the size, hold back a separate evaluation set first, and spend on quality before quantity.
Data, last revised November 7, 2026
LoRA, full fine-tuning and distillation in plain English
LoRA trains a small set of extra weights and leaves the base model untouched, so it is cheaper and easy to swap. Full fine-tuning updates every weight, costs more and can go further. Distillation trains a smaller model to copy a larger one's outputs, trading some quality for much cheaper serving.
Methods, last revised November 7, 2026

Blog