Fine-tuning

What it is

Fine-tuning takes a pretrained model and keeps training it on a smaller, task-specific dataset. OpenAI’s docs describe a loop: write evals, prompt the model with good context, fine-tune only where it helps, and measure again.

How it works

1. Collect high-quality input and output examples of the task.
2. Train a copy of the model on them, through a vendor service or on open weights (for example, models from the Hugging Face Hub).
3. Evaluate the tuned model against the base model on held-out examples before switching.

Where it’s available (checked 2026-10-17):

• OpenAI. OpenAI’s deprecations page says it notified developers on May 7, 2026 that self-serve fine-tuning is winding down. Organizations that had never run fine-tuning lost access on May 7, 2026. Organizations with no fine-tuned-model inference in the past 60 days lost the ability to create jobs on July 2, 2026. From January 6, 2027, active existing customers can’t create new fine-tuning jobs. Fine-tuned models stay available for inference until their base models are deprecated.
• Mistral. Mistral’s fine-tuning docs say the feature is deprecated and no longer actively supported. The page gives no date.
• Google Cloud documents supervised, preference and reinforcement tuning for Gemini models.
• Open-weight models can be tuned on your own infrastructure.

Benefits

• Consistent output format and style without long prompts.
• A small tuned model can replace a large general one on a narrow task, cutting cost.
• Useful for classification and extraction at volume.

Limits and risks

• Needs good data and evaluation. Bad examples make a worse model.
• Facts learned this way go stale. Retrieval is easier to update.
• Ties you to one model version, so you re-tune when it’s retired. Recent vendor shutdowns show that the service itself can go away.