Short answer: for most business tasks, fine-tuning an open-weight model with LoRA or QLoRA costs far more in people time (data preparation and evaluation) than in GPU time. A typical project takes three to eight engineer-weeks end to end. Training compute for a 7–14B parameter model on a few thousand examples is often a small part of the budget. Serving the model is an ongoing cost to plan for. These are typical ranges, not a quote.
Typical effort ranges
| Phase | What happens | Typical effort |
|---|---|---|
| Feasibility and baseline | Measure prompting/RAG performance; decide if fine-tuning is warranted | 1–2 weeks |
| Dataset preparation | Collect, clean, label, deduplicate, split; synthetic data with validation | 1–4 weeks |
| Training experiments | LoRA/QLoRA runs across base models and settings | 0.5–2 weeks |
| Evaluation | Task metrics, regression and safety checks, human review | 1–2 weeks |
| Deployment | Quantisation, serving, monitoring, retraining recipe | 0.5–2 weeks |
GPU compute maths
Training cost ≈ GPU hourly price × hours × number of runs.
As an illustration with stated assumptions: fine-tuning an 8B model with QLoRA on 5,000 examples of around 1,000 tokens each, for three epochs, might take on the order of one to a few hours on a single modern data-centre GPU. If such a GPU rents for a hypothetical $2 per hour, each run costs a few dollars. Running ten to twenty experiments to find good settings still keeps compute in the tens to low hundreds of dollars. Larger models, longer sequences and full fine-tuning increase this substantially. GPU marketplace prices vary by provider and change often, so check current rates.
The larger costs
Data. The quality of the dataset determines the quality of the model. If you have well-labelled historical examples, preparation is quick. If examples must be created by experts, that time dominates the project. Synthetic data can help, but it must be validated.
Evaluation. A trustworthy evaluation set, automated scoring and human review are essential. Without them you can’t tell whether the fine-tuned model is actually better than the base model with a good prompt.
Serving. A fine-tuned model needs to run somewhere:
- Rented GPU, always on. Predictable monthly cost; best for steady traffic.
- Serverless GPU or managed endpoints. Pay per use; good for spiky or low traffic, with cold-start trade-offs.
- Quantised on CPU or small GPU. Smaller models quantised to 4-bit can run on modest hardware for low-volume use.
- Provider-hosted fine-tunes. Some API providers host fine-tuned models for a per-token price.
When fine-tuning pays off
Fine-tuning tends to pay for itself when a high-volume task currently uses an expensive model, when you need consistent output formats or style, or when privacy demands a self-hosted model. If your problem is that the model doesn’t know your facts, start with RAG instead — see our RAG vs fine-tuning comparison.
Getting an estimate
Send a description of the task, a sample of the data you have and your expected volumes. A one- to two-week feasibility study establishes the baseline and gives a fixed price for the full project.