What is fine-tuning an LLM?

Fine-tuning takes a model that has already finished training and adapts it to one task, usually with LoRA or QLoRA, which freeze the base model and attach a small add-on rather than rewriting the whole thing. What it changes is skill, output format and taste. It is not how you teach a model new facts.

What is the difference between fine-tuning and RAG?

They solve different problems. RAG opens your documents at answer time, so it suits new knowledge and knowledge that changes often. Fine-tuning bakes behaviour into the model itself, so it suits skill and format. Push knowledge into the weights and that knowledge is frozen there, going stale as the world moves.

How many stages does training a model have?

Four. Training from scratch, continued training on domain documents, post-training that makes it obedient and sharp, then fine-tuning it to one task. The first three are what create intelligence. The fourth only adjusts behaviour. The line between the third and the fourth is what makes the decision possible.

When is fine-tuning worth it?

The default is that it is not, because the base model is already good and the gain is usually small. Three reasons still hold in 2026: cutting cost per call, forcing output into one consistent schema, and shrinking a model while keeping the same job. And before starting you need an eval set of your own, or there is no way to know whether anything improved.

When Fine-Tuning an LLM Is Worth It: Four Stages, and the Line Down the Middle

When a model answers off-target for your work, the first thought is usually the same one: we need to train it on our data.

This block of AI Update Bangkok 2026 is the one that makes you rethink that, because the answer from Dr. Jim (Jimmy Tejasen), who hosts the event, is short and cuts against the expectation. Most of the time it is not worth it. And the reason is not that it is expensive or hard. It is that training solves a different problem from the one most people actually have.

Following part two, the map of the whole day and part three, on compressing models, this part lands on training, which is the block he built as decision rules rather than product news.

Terms, defined once, here

LLM a large language model, the thing that reads text and writes text back; the engine underneath the assistants people use
base model a model fresh out of the first stage; all it can do is guess the next word, and it cannot hold a conversation
instruct model the one people actually use: a base model that has been taught to follow instructions
fine-tune adapting a finished model to one task
LoRA a way of fine-tuning that freezes the base model whole and attaches a small add-on instead of rewriting it; QLoRA does the same on a base compressed to 4 bits, so it fits a smaller card
RAG letting a model answer from your documents by opening them at question time rather than baking them in
eval a set of questions with known-correct answers that you build yourself, to measure whether a model got better or worse at your work

Part 1The training conveyor has four stages

The phrase training a model gets used for so many things that it has stopped meaning anything, because it covers four different jobs whose price and outcome are worlds apart. So he starts by laying the conveyor out.

Stage	What it does	What comes out
1 From scratch	Read an enormous amount of text and predict the next token	A base model that cannot yet hold a conversation
2 Continued, domain	Keep training on the documents of one field	The same model, now fluent in that field's terms
3 Made obedient	Teach it to follow instructions, then sharpen it with feedback	The model people actually use
4 Adapted to one task	Adjust the finished model for a single job	The same model with different habits

The first three belong to whoever owns the model, using datacentre hardware and months of time. Stage four is the only one an ordinary team can do, and it is the only one people mean when they say they are going to train a model.

Part 2The line drawn down the middle

The single sentence that makes the whole subject decidable sits at the seam between stage three and stage four.

Stages 1 to 3 create intelligence. Stage 4 adjusts behaviour.

Which means if what you want is on the intelligence side, fine-tuning cannot give it to you, no matter how good your data preparation is. It is not the tool for that side.

Once that line is clear, two questions remain: what to teach, and how to teach it.

What to teach ranges from feeding it correct question-and-answer pairs directly, through to showing it two candidate answers and saying which one is better. That second kind teaches taste in a way that plain examples cannot.

How to teach it has three levels: rewrite the whole model, freeze the base and attach an add-on, or attach an add-on to a base that has already been compressed. The last two exist so this work fits the machine you already own, which is why people can start experimenting without renting a datacentre.

Worth attaching: the ease of the procedure does not change the answer about whether it is worth doing. What it changes is the price of trying.

Part 3RAG is not training

The point he pressed hardest in this block is also the one people confuse most.

The problem you have	The right fix
The model knows nothing about the inside of your company	Open the documents at answer time
The material changes every week	Open the documents at answer time
Right answers, unstable format, nothing downstream can consume it	Adjust behaviour by fine-tuning
The voice is not yours	Adjust behaviour by fine-tuning

The split is not a matter of preference. Knowledge pushed into a model's weights is frozen there. The day the material changes, the model keeps answering confidently from the old version, and the fix is another full training round.

Documents opened at answer time are fixed by editing a file.

Part 4So when is it worth it

His starting answer is that it is not, and the reason is stated plainly: the base model is already good, so the increment from training is usually smaller than people expect.

More interesting is that he says the role of fine-tuning has changed in 2026. It has gone from a tool for making a model smarter to a tool for cutting cost and forcing format. Optimising the instructions now competes with the more elaborate training methods, and beats them on plenty of jobs.

What is left for training to do comes down to three things.

Cut cost per call. Teach a small model to do the same job instead of paying for a large one every time.
Force the output shape into one consistent structure, so downstream systems can consume it without a layer of patching.
Shrink it so it fits the machine you already own, while holding the same job.

Notice that not one of the three says smarter. All three are about cost and certainty.

Put the three conditions next to the default answer and you get an order you can actually decide with: polish the instructions first; if the problem is knowledge, open the documents at answer time; then think about training, and only once you know how you will measure it.

Part 5Which part to read next

That closes the training block. One block is left, and it is the one about money.

Part 1 Physical AI, and whether a mistake can be undone
Part 2 The map of the whole day, and four questions before believing a number
Part 3 Compressing models, and the quality cost you cannot see
Part 5 Which layer of the AI stack collects the money, and why the wrapper has the power

Source

Everything here comes from AI Update Bangkok 2026, held on 21 August 2026 at The Cloud Bangkok by Jimmy Tejasen. This part covers the session on open-weight models and training.

Event livestream: watch the recording
Jimmy Tejasen's work: aiserver.in.th, news and reviews of AI server hardware, and agentic-press.com, a publishing house for AI books

The figures here come from what was said on stage, not from vendor spec sheets. Use them for ratios and for asking better questions, not as product specifications. Productize has not reproduced anything shown at the event.

AI Update Bangkok 2026, Part 4