Cost
How to estimate what a fine-tuning run will cost
How much does it cost to fine-tune an open-weight model?
Cost is the hourly rate of the GPU class multiplied by wall time. Wall time is driven by total training steps, which is dataset size divided by effective batch size, times epochs. A LoRA run on an 8B model over a few thousand examples is usually well under an hour.
The arithmetic
Two numbers multiply: the hourly rate of the class and how long the run takes. The rate is published. The time is the part worth estimating properly.
Total steps is roughly your example count divided by the effective batch size, multiplied by epochs. Throughput in steps per second depends on model size, sequence length and whether memory-saving techniques are in use — each of which trades time for memory.
The cheapest class is often not the cheapest run
This is the most common budgeting error. A faster class at a higher hourly rate can finish sooner and cost less overall, particularly when the alternative is four-bit training on a smaller card, which is meaningfully slower per step.
Compare total run cost rather than hourly rate. Where a run fits half-precision on a dearer class and needs quantisation on a cheaper one, the dearer class frequently wins.
Where budgets actually go wrong
- Sequence length set to the model maximum by default, paying for headroom the data never uses.
- Too many epochs. Fine-tuning wants one to three; runs configured for ten are paying to overfit.
- An endpoint left warm after an experiment. Serving is where lifetime cost accumulates, and an idle warm endpoint is the single most common source of waste.
- Retraining after a failure that dataset validation would have caught for nothing.
Training is the small number
For a model that goes into production, training is a handful of runs and inference is every request forever. A model twice as large costs roughly twice as much on every request for the rest of its deployment.
This is the strongest argument for starting at the bottom of the size ladder and moving up only when the gate says you must. The saving compounds in a way the training saving does not.
Forecast before you commit
A forecast before submission turns this from an estimate into a number, and the cost is then reserved against your limit before compute is leased. A run that would breach the ceiling does not start.
Watching the gap between forecast and actual over several runs is also the fastest way to calibrate your own intuition about which levers matter.
Frequently asked questions
Do I pay while waiting for capacity?
No. Metering starts when the job is on the GPU. Queue time and teardown are not billed.
What is the minimum billable amount?
Compute is metered per GPU-second above a one-minute floor, so a very short job is billed as one minute rather than rounded up to an hour.
Terms used here
More on cost
Last verified 6 August 2026.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.