Fine-tuning is a small nudge on top of a huge model
A language model arrives at you already trained. Pretraining read a corpus of trillions of tokens; Meta reported more than 15 trillion for Llama 3. Post-training then coached it into an assistant. What is an LLM covers both stages.
Fine-tuning is a third, much smaller stage that you run. You give the model a few hundred to a few thousand examples of the exact input and output you want, and train it a little more on those. OpenAI's fine-tuning guide says clear improvements typically show from 50 to 100 examples. The parameters shift slightly toward producing your kind of answer by default.
The scale difference is the whole story. Fifteen trillion tokens against a few hundred examples: the fine-tune can't add meaningful knowledge, because there isn't enough of it to move the needle on what the model knows. What it can do is move the needle on how the model responds, because a consistent format repeated a few hundred times is exactly the kind of pattern a small nudge captures.
Fine-tuning changes how a model answers, never what it knows
The line to draw is behaviour against knowledge.
| You want the model to... | Fine-tuning | What actually works |
|---|---|---|
| Always answer in your house style or tone | Yes, this is the core use | Fine-tune, or a strong system prompt with examples |
| Always return a specific format (JSON, a fixed template) | Yes | Fine-tune for reliability at scale; a prompt for occasional use |
| Follow a long list of rules without being reminded | Yes | Fine-tune; the rules move from the prompt into the weights |
| Know your product, prices, policies | No | Retrieval: hand it the documents at answer time |
| Know anything that changed after training | No | Web search or retrieval |
| Reason better on hard problems | No | A reasoning model or a bigger model |
The failure people hit is the fourth row. They fine-tune on their help centre, expecting a model that knows their product. They get a model that writes in the help centre's voice and invents the product facts fluently, because the training taught it the shape of the answers and not the substance. The knowledge job belongs to retrieval, and what is RAG explains how that works.
Why fine-tuning is done on cheap models on purpose
Nobody fine-tunes the most expensive model in a family for a narrow job. The point is to get a small, cheap model to do one thing as well as the frontier model does it with a long prompt, and then run the cheap one millions of times.
The numbers above each end are Artificial Analysis Intelligence Index scores, read on 2026-09-17, and the line is the price gap per million tokens. A general-purpose Llama 4 Maverick scores 9 against Claude Sonnet 5 at 38; the bet of a fine-tune is that on one narrow task those 29 points can be closed for a 127-fold saving. At OpenRouter list prices fetched on 2026-08-20, a standard answer of 1,000 tokens in and 500 out costs $0.0006 on Maverick and $0.007 on Sonnet 5, per the AI model cost index. A support classifier that runs 100,000 times a month costs $60 on the first and $700 on the second. If the tuned Maverick matches Sonnet on that one task, the tune pays for itself in the first month. That arithmetic, not model quality in general, is what fine-tuning is for.
The method that made this cheap is LoRA, published in 2021, which trains a small set of added parameters rather than the whole model. With its 4-bit successor QLoRA (Dettmers et al., 2023), a 7-billion-parameter model tunes on a single 24 GB consumer graphics card. Full fine-tuning of the same model means updating every parameter across several datacenter GPUs.
A prompt, retrieval or a different model solves most cases before any training
Fine-tuning is the last option, not the first, because each of these solves the common cases for a fraction of the effort and can be changed in a minute.
- A better prompt with examples. Three examples of the output you want inside the prompt gets most of the format and tone benefit of a fine-tune with none of the training. Prompt engineering for beginners covers the pattern.
- Retrieval. If the gap is knowledge, no amount of tuning closes it. Upload the documents or turn on search.
- A different model. Models differ in default voice. A model that already writes the way you want beats tuning one that does not; a multi-model workspace makes this a one-click test.
Reach for fine-tuning when the prompt has grown to a page of rules the model keeps forgetting, when the job runs at a volume where prompt length is real money, or when you need a format returned reliably ten thousand times without a single deviation. Those are real needs. They are also rare.
Can you fine-tune inside a chat product?
Generally no, and Whizi is no exception. Consumer chat products, Whizi included, don't offer fine-tuning; it's a developer feature on provider APIs and open-weight models. The nearest thing inside a chat app is standing instructions. In Whizi those live on a project: per-project custom instructions apply to every conversation in that project, which is how you get consistent tone and format across a body of work without training anything. How to use Projects explains the setup.
That's the right trade for most people. A project instruction is a prompt you write once and edit whenever the rules change. A fine-tune is a model you retrain whenever the rules change. For everyone who isn't running a narrow task at scale, the prompt wins on cost, speed of change and the fact that you can read it.
Fine-tune when the same narrow job runs at a volume where a cheap model paying for itself matters. Everywhere else, write the instruction down.
- Decide whether the gap is behaviour or knowledge; fine-tuning only fixes the first
- Try a prompt with three examples before any training
- Use retrieval or web search for facts, prices and anything that changes
- Test a different model first; one may already write the way you want
- Fine-tune a cheap model for a narrow, high-volume job, and measure it against the frontier model on that job
Frequently asked questions
What does fine-tuning an AI model mean?
It means training an already-trained model a little more on your own examples, typically a few hundred to a few thousand input-output pairs, so it adopts a style, format or set of rules by default. It's a small adjustment on top of pretraining that read trillions of tokens, which is why it shapes behaviour rather than adding knowledge.
Fine-tuning vs RAG: which should I use?
Fine-tuning for how the model answers: tone, format, following a long rule set. RAG for what it knows: your documents and anything that changes. They solve different problems and production systems often use both. If your gap is facts, fine-tuning won't close it; the model will sound right and be wrong.
How many examples do you need to fine-tune a model?
OpenAI's fine-tuning guide says clear improvements typically show from 50 to 100 well-made examples, and that quality matters more than count. A few hundred is a common working size; thousands help for tasks with many distinct cases.
How much does fine-tuning cost?
The training run is usually the small part; the point is the per-answer saving afterwards. At OpenRouter list prices fetched 2026-08-20, a standard answer is $0.0006 on Llama 4 Maverick and $0.007 on Claude Sonnet 5, so a job that runs 100,000 times a month costs $60 against $700. A fine-tune that lets the cheap model match the expensive one on that job pays back in the first month.
Can I fine-tune ChatGPT or Claude?
Not the consumer apps. Fine-tuning is offered on provider APIs for some models and on open-weight models you run yourself. Inside a chat product the equivalent is a standing instruction: ChatGPT's custom instructions, or a project-level instruction in Whizi that applies to every conversation in that project.
