Provider pages and project proposals use these three terms as if everyone knows them. They describe different ways of changing a model, with different costs and different things you end up owning. Here is what each one does, without the math.
What a model’s weights are, briefly
A language model is a very large set of numbers (weights) learned during pretraining. Fine-tuning means adjusting the model so it behaves the way your examples show. The question is how much of it you adjust, and whether you keep the same model at all.
LoRA: train a small add-on, leave the base alone
LoRA (low-rank adaptation) was introduced by Hu and colleagues in 2021. Instead of changing the model’s own weights, it freezes them and trains a small set of extra weights that sit alongside certain layers. At answer time the add-on nudges the base model’s behavior toward your examples.
Why teams like it:
- Cheaper training. Far fewer numbers are being trained, so jobs need less GPU memory and time. Fireworks AI, for example, publishes LoRA SFT at half its full SFT rate in every model-size tier.
- Small files. The trained adapter is a fraction of the size of the model, so it is easy to store, version and swap.
- Many tasks, one base. Some serving setups can load several adapters over a single base model, which helps when you have many narrow tasks.
The limit: LoRA moves the model a shorter distance. For most format, style and narrow-task work it is enough. For large shifts in behavior, it can fall short.
Full fine-tuning: update every weight
Full fine-tuning trains all of the model’s weights on your data. The result is a complete new copy of the model.
What that buys and costs:
- More room to change. It can learn larger shifts than an adapter, which matters for broad or unusual tasks.
- Higher training cost. On Fireworks AI’s published table, full SFT costs twice LoRA SFT at every size, from $1.00 per million training tokens for models up to 16B to $20.00 for models over 300B (read 7 Nov 2026).
- Bigger to host. You serve a whole model, not an add-on, so you need the same hardware the base model needs, for your copy alone.
- More risk of forgetting. Updating every weight on a narrow set can erode general abilities. The evaluation set needs a few general-purpose checks, not only task cases.
Distillation: teach a small model to copy a big one
Distillation is older than either method above; Hinton, Vinyals and Dean described the modern form in 2015. A large “teacher” model produces outputs for many inputs, and a smaller “student” model is trained to reproduce them.
The point is serving cost. If a large model does your task well but costs too much per request at your volume, a distilled small model can do the same narrow task for much less per request, often faster. It will usually give up some quality, especially at the edges of the task.
Two practical notes. First, distillation needs a lot of teacher outputs, so its data cost is in running the teacher, not paying people. Second, check the teacher model’s terms: some providers restrict using their outputs to train other models. AWS Bedrock lists distillation among its customization options, though its distillation price table was not captured when we read the page.
Where SFT, DPO and reinforcement fine-tuning fit
These are different questions from LoRA versus full. They describe what the model learns from:
- SFT (supervised fine-tuning): pairs of input and ideal output.
- DPO (direct preference optimization): pairs of a better and a worse answer to the same input, introduced by Rafailov and colleagues in 2023.
- Reinforcement fine-tuning: a grader scores the model’s attempts and training rewards higher scores.
You can combine them: LoRA with SFT, full fine-tuning with DPO, and so on. Pricing pages are laid out this way, as the table below shows.
| SFT | DPO | |
|---|---|---|
| LoRA, models up to 16B | $0.50 | $1.00 |
| Full, models up to 16B | $1.00 | $2.00 |
| LoRA, 16.1B to 80B | $3.00 | $6.00 |
| Full, 16.1B to 80B | $6.00 | $12.00 |
Fireworks AI list prices in dollars per million training tokens, copied 7 Nov 2026; not a quote. A provider on this site may pay us after an introduction, which has no bearing on numbers we reproduce.
Which one should you pick?
| Situation | Usually |
|---|---|
| Narrow task, format or style, first attempt | LoRA with SFT |
| Several narrow tasks on one base model | LoRA adapters, one per task |
| Large behavior change, plenty of clean data | Full fine-tuning |
| A large model works but costs too much per request | Distillation into a smaller model |
| You can rank answers but not write ideal ones | DPO after an SFT pass |
| Success can be checked by a program | Reinforcement fine-tuning |
Start with LoRA unless you have evidence it falls short. It is the cheapest way to find out whether fine-tuning helps at all, and the adapter can be thrown away if it does not. The fine-tuning price tables show how each platform bills these methods, and the LLM fine-tuning services page sets out the decision from the top.
What you own at the end
Ask this before choosing. With LoRA you own an adapter that only works with its base model. With full fine-tuning you own (or rent the serving of) a whole model. With distillation you own a smaller model that is yours to host. Whether you can download any of these depends on the platform, so get it in writing.
The method is a cost decision as much as a technical one, and the cheapest one that passes your evaluation is the right one.
Sources for this post
- Hu et al., "LoRA: Low-Rank Adaptation of Large Language Models", arXiv 2106.09685 (2021)
- Hinton, Vinyals and Dean, "Distilling the Knowledge in a Neural Network", arXiv 1503.02531 (2015)
- Rafailov et al., "Direct Preference Optimization", arXiv 2305.18290 (2023)
- Fireworks AI pricing (LoRA and full SFT and DPO by model size), read 7 Nov 2026
- Amazon Bedrock pricing (customization tables, partial), read 7 Nov 2026