Skip to content
HomeBriefFor youTrackersTopicsCompaniesGlossaryNewsletterSecurityHealth & medicineFinance & marketsRoboticsModels & researchPolicy & regulationEnterprise & businessDeveloper tools
Explainer · beginner · 3 min read

What is fine-tuning? How models get specialised, when it beats prompting and retrieval, and why a fine-tuned checkpoint is now a supply-chain risk

Fine-tuning takes a trained model and keeps training it on your examples. Here is what actually changes inside the model, the three cheaper alternatives to try first, what it costs in 2026, and the provenance checks that this week's Microsoft research made mandatory.

Updated 2026-11-15 · For: Beginners Developers ML engineers Executives Security

A language model is a large set of numbers, its weights, learned by predicting text across trillions of tokens. Fine-tuning means taking those weights and continuing to train them on a much smaller set of examples that show the behaviour you want: your company's tone, a medical coding scheme, a telecom's network vocabulary, a particular output format. The model keeps almost everything it knew and shifts a little toward your data. It is the step that turns a general base model into the chat assistant you talk to, and the step that turns a chat assistant into a specialist.

What changes inside the model

Pretraining sets the weights by exposure to the open internet; fine-tuning nudges them with a few thousand to a few million curated examples. Full fine-tuning updates every weight and needs the same hardware as training. Parameter-efficient methods such as LoRA freeze the original weights and learn a small adapter, a few tens of megabytes, that is added on top at inference time; AT&T's telecom models, built on Google's Gemma 4, and most of the fine-tunes on public hubs are made this way. Reinforcement fine-tuning goes further and trains on a reward signal rather than fixed examples, which is how reasoning and agentic behaviour are taught. In every case the result is a new checkpoint whose weights differ from the base model's.

When to fine-tune, and when not to

Three cheaper things usually come first. Prompting with a good system prompt and examples handles most tone and format problems. Retrieval-augmented generation gives the model your documents at query time and is the right tool when the problem is missing knowledge, because knowledge changes and a fine-tuned model does not. Picking a better or larger model often beats both. Fine-tuning earns its cost when the behaviour is hard to describe but easy to show, when you need a smaller model to match a larger one on a narrow task to cut inference bills, when latency rules out long prompts, or when the vocabulary is genuinely foreign to general models, as in telecom operations or clinical notes. It does not reliably add facts, and it can quietly erase safety behaviour the base model had.

What it costs

For open-weight models a LoRA run on a mid-sized model costs tens to a few hundred dollars of GPU time; the real cost is building the dataset. Hosted services charge per training token and then a premium at inference: OpenAI's published fine-tuning rates for its remaining supported models are $25 per million training tokens for GPT-4.1 with inference at $3 and $12 per million, roughly three to four times the base price, and the company has closed fine-tuning to new users as it moves customers to prompting and retrieval on newer models. Thinking Machines' Tinker and similar platforms sell fine-tuning of large open models as a managed service. Factor in evaluation: a fine-tune without a held-out test set is a guess.

The new risk: a checkpoint is code you did not write

Every fine-tune, merge, quantised build and adapter on a public hub is a modified copy of someone else's model, redistributed under a familiar name. Until this month that was mostly a quality problem. On November 10 Microsoft Research showed that it is a security problem: by editing fewer than ten dormant neurons, with no training data at all, they planted a backdoor that made a coding agent emit an attacker's shell command in 98.8% of triggered sessions while passing ordinary benchmarks. Honest fine-tuning spreads small changes across the network; an implant concentrates a large change in a few weight vectors. The only reliable way to tell them apart is to compare a checkpoint against the publisher's original, which means you must know what the original is. Model provenance, pinned hashes, a private registry and a weight diff before deployment are now part of fine-tuning hygiene, on the receiving end as much as the producing end.

How to do it well

Start with the smallest model that could plausibly work and a few hundred high-quality examples; more data of lower quality rarely helps. Hold out a test set the model never sees and measure against the un-tuned model and against a strong prompt, so you know the fine-tune paid for itself. Re-run your safety and injection evaluations on the tuned model, because fine-tuning can undo refusals. Record the base model's hash, the dataset version and the training configuration with the checkpoint. If you distribute the result, publish the base hash too; if you consume someone else's, diff it before it touches an agent with tool access.

Common questions

Does fine-tuning teach a model new facts?

Not reliably. It shifts style, format and behaviour far more than it adds knowledge, and facts baked into weights go stale. Use retrieval for knowledge and fine-tuning for behaviour; the two combine well.

Can I fine-tune a closed model like Claude or Gemini?

Only through the vendor's own service where one exists, and the vendor keeps the weights. Several vendors have narrowed or closed fine-tuning as newer models handle more through prompting; check the current pricing page before planning on it.

How do I know a downloaded fine-tune is safe?

You cannot prove it, but you can check it: verify the publisher, pin the hash, and compare the weights with the declared base model. Large changes concentrated in a few vectors are a red flag. Run it in a sandbox with minimal permissions until it has earned trust.

Get the brief in your inbox

One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.

Published daily by AI Briefing. We never share your address.
Like this? Get tomorrow's brief by email, free.