How do you fine-tune a small open model yourself?

Fine-tuning means continuing to train an existing model on your own examples so it adopts a habit: a fixed format, a house style, a way of sorting requests. It does not teach a model new facts reliably (for that, giving it documents at question time works better), but it is excellent for teaching it how to answer.

In October 2026, a small open-weight model can be tuned on a free cloud notebook or a gaming graphics card, mostly thanks to two techniques. LoRA freezes the original model and trains a small set of extra weights, called an adapter, beside it; our August guide walks through a LoRA run with Hugging Face's libraries. QLoRA goes further by storing the frozen model in 4-bit form while the adapter trains.

Why QLoRA changed the budget

The QLoRA paper, published in May 2023, showed that a 65-billion-parameter model could be fine-tuned on a single 48 GB graphics card while matching the quality of full 16-bit fine-tuning. Its best model reached 99.3% of ChatGPT's score on the benchmark the authors used after 24 hours on one GPU. Just as important for beginners, the authors reported top results from training on a small but carefully chosen dataset.

The tools have since made this routine. Unsloth, a popular open fine-tuning library with a desktop app, now trains on NVIDIA, AMD and Intel graphics cards and on Macs, and offers free notebooks for Google Colab and Kaggle. Its documentation lists the minimum graphics memory (VRAM) needed:

These are minimums; longer examples need more. An 8B model with QLoRA fits on a 16 GB card such as the Nvidia T4, a common choice for free and low-cost notebooks.

A worked run, step by step

Say you want a model that turns messy customer messages into a tidy support ticket with four fields.

1. Choose a base model. Start with an instruction-tuned open model of 3 to 8 billion parameters under a licence that allows your use. Small models train fast, so you can iterate.
2. Write the examples. Put each one in the chat format the tool expects: the customer message as the user turn, the ideal ticket as the assistant turn. Consistency matters more than volume: every example should follow exactly the format you want back.
3. Set some aside. Keep 10 to 20% of examples out of training. They are your exam.
4. Train. Load the base model in 4-bit, attach a LoRA adapter, and train for one to three passes over the data (epochs); Unsloth's guide recommends staying in that range to avoid memorizing the examples. If you run out of memory, lower the batch size to 1 or 2 before anything else.
5. Test against the original. Run both the base model and the tuned one on the held-out examples and score them the same way: does each ticket have all four fields, in the right format, with the right category?
6. Keep or discard. If the tuned model is clearly better, save the adapter, a small file next to the base model, and load it with the base model. If not, improve the data before you train again.

How much data and money does it take?

Data. Unsloth's dataset guide suggests a bare minimum of about 100 examples to see a reasonable result, with more than 1,000 preferable; beyond that, more good data usually helps. For a narrow format task, a few hundred carefully checked examples often go a long way. The real cost is your time: writing, reviewing and correcting examples usually takes far longer than training. You can ask a larger model to draft extra examples, but check them by hand and read the article on distillation in this series first, because the terms of some services restrict using their outputs to train other models.

Money. If you already own a recent graphics card with enough memory from the table above, the cost is electricity. The free Colab and Kaggle notebooks cost nothing but come with session limits. To rent, Hugging Face Jobs, which bills by the minute, lists hourly prices in October 2026 such as:

So if a QLoRA run on an 8B model takes two hours on an L4, the bill is $1.60. Even ten experimental runs while you refine the data stay under $20. Set a timeout on every job so a forgotten run cannot keep billing. The expensive part of fine-tuning in 2026 is not the computer; it is building the examples and the test that tell you whether it worked.