Which fine-tuning method should I use?
Three ways to adapt a model, separated by roughly an order of magnitude in memory each. For most people the choice is made by the card, not by preference.
The short answer
Computed from this tool’s default settings — your hardware and the rest as most people start. Change them below for your own case.
RTX 4090 QLoRA-fine-tunes up to a 27B model. Full fine-tuning is effectively server-only above about 3B, because optimiser state alone is roughly 12 bytes per parameter.
LoRA vs QLoRA vs Full Fine-Tune
Memory, fidelity and hardware for all three, at every model size.
RTX 4090 QLoRA-fine-tunes up to a 27B model. Full fine-tuning is effectively server-only above about 3B, because optimiser state alone is roughly 12 bytes per parameter.
| Size | QLoRA | LoRA | Full | On your card |
|---|---|---|---|---|
| 1B | 1.20 GB | 2.70 GB | 16.0 GB | LoRA or QLoRA |
| 3B | 3.50 GB | 8.00 GB | 48.0 GB | LoRA or QLoRA |
| 7B | 5.00 GB | 19.0 GB | 112.0 GB | LoRA or QLoRA |
| 8B | 6.00 GB | 22.0 GB | 128.0 GB | LoRA or QLoRA |
| 14B | 8.50 GB | 33.0 GB | 224.0 GB | QLoRA only |
| 27B | 22.0 GB | 64.0 GB | 432.0 GB | QLoRA only |
| 32B | 26.0 GB | 76.0 GB | 512.0 GB | Neither |
| 70B | 41.0 GB | 164.0 GB | 1,120.0 GB | Neither |
| 405B | 237.0 GB | 950.0 GB | 6,480.0 GB | Neither |
What each setting changes
Every input moves the result for a reason. This is what each one does and where to find the value for your own machine.
| Setting | Default | What it changes |
|---|---|---|
| Your hardware | RTX 4090 · 24 GB | The machine the model runs on. Usable memory decides what fits and memory bandwidth decides how fast it runs, so this moves every figure below. |
Real answers across your hardware
The same calculation run at a range of settings, with everything else left at its default. These are computed by the tool itself, not written by hand.
| Your hardware | Largest you can QLoRA | Your memory | Method to prefer |
|---|---|---|---|
| NVIDIA B200 · 180 GB | 70B | 165.6 GB | QLoRA |
| Cerebras WSE-3 · 44 GB on-chip SRAM | 70B | 44.0 GB | QLoRA |
| Raspberry Pi 5 · 16 GB | 14B | 9.60 GB | QLoRA |
| RTX 4080 SUPER · 16 GB | 14B | 14.7 GB | QLoRA |
| RTX 5070 · 12 GB | 14B | 11.0 GB | QLoRA |
NVIDIA B200 QLoRA-fine-tunes up to a 70B model. Full fine-tuning is effectively server-only above about 3B, because optimiser state alone is roughly 12 bytes per parameter.
How this is calculated
No lookup tables and no invented constants. Here is the arithmetic, so you can check it against your own numbers.
QLoRA quantises the base model to 4-bit and trains small adapters on top, so the card that runs a model can usually train it too. LoRA keeps the base at 16-bit and trains the same adapters, costing roughly three to four times more memory. A full fine-tune updates every weight and carries gradients plus optimiser state — about 16 bytes per parameter.
The QLoRA and LoRA figures are Unsloth’s published requirements, interpolated between the sizes they list rather than extrapolated from a formula. They assume a short sequence length and a small batch; longer sequences and larger batches push the real number up.
These are well-founded engineering estimates, not benchmark results. Your quantisation, runtime and context length all move the real number, and usable memory is an assumption rather than a specification. See the full methodology for every assumption behind these figures.
Who this is for
The three situations that bring people to this calculation.
Deciding an approach
See all three against your hardware at once.
Budgeting a rental
Know what each method costs in GPU hours.
Learning
Understand where the memory actually goes.
How to use this calculator
Four steps, no account, nothing leaves your browser.
Set your hardware
Start at the top of the panel. Every figure recalculates as you change it — there is no submit button, because watching the number move is the point.
Adjust the rest to match your setup
0 further settings: . Defaults are the common case, so change only what differs for you.
Read the headline, then the table
The large figure answers the question. The table underneath shows how the answer changes across nearby settings, which is usually where the decision actually gets made.
Check it against the method
The arithmetic is written out above. If a number looks wrong for your hardware, the assumptions are the first place to look — usable memory and quantisation are the two that vary most.
Which fine-tuning method should I use: common questions
The questions people ask about this, answered without hedging.
What is the difference between LoRA and QLoRA?
Both train small adapter matrices and leave the base frozen. QLoRA additionally quantises that frozen base to 4-bit, cutting memory by roughly three to four times for a small quality cost.
When is a full fine-tune worth it?
Rarely outside a lab. It updates every weight and needs about 16 bytes per parameter for weights, gradients and optimiser state, which is server territory above 3B. Adapters get most of the benefit for a fraction of the cost.
Can I merge a LoRA back into the model?
Yes. Merging folds the adapter into the base so you can serve a single model with no runtime overhead. Keep the unmerged adapter too — it is a few megabytes and lets you rebase onto a newer version later.
Why is full fine-tuning 16 bytes per parameter?
Two bytes of fp16 weights, two of gradients, and twelve for Adam’s two moments plus an fp32 master copy. The optimiser state, not the model, is what dominates.
Related local AI tools
All 50 run on the same arithmetic, so answers across them agree.
More on fine-tuning
Will it fit
Context
Speed
Quantisation
Runtime & errors
Cost
Searches this page answers
Different ways of asking the same question, all resolved above.
Now find the models that fit
Sizing is only half the problem. Model Radar takes your hardware and shows which models actually run on it, ranked by what they are good at — the same arithmetic as this page, applied to every model worth running.