AI Fine-tuning

AI fine-tuning is additional training of an existing foundation model on a focused dataset, so the model's weights change to suit a task: a house writing style, a fixed output format, a classification scheme or a domain's vocabulary. It costs far less than training a model from scratch and produces a model that behaves differently without long instructions in every prompt.

The main methods are supervised fine-tuning (SFT) on pairs of inputs and ideal outputs; preference tuning such as DPO, which learns from pairs of better and worse responses; reinforcement fine-tuning, which scores the model's attempts with a grader and suits reasoning tasks; and parameter-efficient methods such as LoRA and QLoRA, which train small adapter weights instead of the whole model and cut hardware needs sharply. Distillation, training a smaller model on a larger model's outputs, is one of the most common uses of SFT. These sit alongside RLHF, which labs use during their own post-training to align models.

Hosted fine-tuning of frontier models has narrowed. OpenAI's documentation says it is winding down its fine-tuning platform: it no longer accepts new users, existing users can create jobs for the coming months, and fine-tuned models stay available until their base models are deprecated. The models it supports for fine-tuning are older ones such as GPT-4.1 for SFT and DPO and o4-mini for reinforcement fine-tuning; GPT-6 Luna, for example, does not support it. Google offers supervised tuning for Gemini models including Gemini 3.5 Flash and the Gemini 2.5 family on Google Cloud. Amazon Bedrock lists Claude 3 Haiku as its only fine-tunable Claude model. Most new fine-tuning work now happens on open-weight models such as Qwen, Gemma, gpt-oss and DeepSeek, which teams can tune and host themselves.

Fine-tuning changes behavior; it is a poor way to add facts that change. For current information, retrieval-augmented generation and tool use are cheaper to update and can cite their sources. A common pattern combines both: a small tuned model handles formatting, routing or classification, while retrieval supplies the facts. Frontier models with million-token context windows and strong instruction following have also removed many old reasons to fine-tune, since a detailed system prompt with worked examples often gets close enough at no training cost.

For GEO, fine-tuning is context more than a lever. Brands cannot fine-tune ChatGPT, Gemini or Claude, so what those assistants say about a brand comes from pretraining data and live retrieval, not from any tuning the brand does. Fine-tuning touches visibility mainly through vertical tools, such as legal research, clinical assistants or shopping bots, built on tuned open-weight models with their own retrieval indexes. Being present in the sources those tools retrieve from, and in the AI training data of future base models, is what reaches them.