Serverless GPU platform
Onrup vs Modal
Should I use Onrup or Modal?
Modal is the better choice if your training loop is custom and you want to own the code. Onrup is the better choice if your job is a standard fine-tune and you would rather not write, maintain and debug the training harness at all.
Serverless compute for arbitrary Python, billed by the second. Not a fine-tuning product — a substrate you build one on.
Where they differ
| Onrup | Modal | |
|---|---|---|
| What you write | A dataset and a template choice | The training code itself |
| Flexibility | Within validated templates | Arbitrary Python |
| H100 rate | $2.59 per GPU-hour training | $3.95 per GPU-hour |
| Billing granularity | Per GPU-second, one-minute floor | Per second |
| Evaluation gate | Built in and blocking | You build it |
On comparable hardware
Modal publishes $3.95 per H100 80GB hour. Our serving rate for the same class is $2.99, which makes theirs 1.3× the rate. Total cost is the rate multiplied by wall time, so this ratio is the starting point of a comparison rather than the end of one.
Priced per second of GPU time: H100 at $0.001097/sec, which is $3.95 per GPU-hour. A100 80GB is $2.50 and L40S $1.95 per hour. There is no fine-tuning product and therefore no per-token training rate; you write the training code yourself. Source, checked 2026-08-06.
The longer answer
This is not really a like-for-like comparison and it is worth saying so. Modal is a substrate: serverless compute that will run whatever Python you give it, extremely well. Onrup is an application built on that idea for one specific job. Comparing them is comparing a language to a framework.
If your fine-tune is standard — a supported base model, a supervised or preference objective, a conventional dataset — then the training harness is undifferentiated work. Somebody has to write the checkpointing, the resumption, the metric logging, the cost tracking and the teardown, and none of it makes your model better.
If your job is not standard, that same harness is a cage. A custom loss, an unusual data pipeline, a multi-stage job that is only partly training — Modal will run all of it and a template-driven platform will not. There is no version of this comparison where we win on flexibility.
The hourly rate difference is real but should not drive the decision. The larger cost in the Modal case is engineering time, and the larger cost in ours is the runs you cannot express.
Where Modal wins
Flexibility, and it is not close. If your workload is not a standard fine-tune — a custom loss, an unusual data pipeline, a multi-stage job that is only partly training — Modal will run it and a template-driven platform will not. Their per-second billing is also genuinely excellent.
Choose them if
- Your training loop is custom and you want to keep it that way
- You already have ML engineers who would rather own the code
- You need one substrate for training, inference and unrelated batch work
On ownership
Weights on Onrup are downloadable from every finished run and publishable to a model hub in one call. On Modal: Your code, your weights, your storage.
Frequently asked questions
Can I bring my own training code to Onrup?
No. Runs execute within validated templates, which is what makes cost forecasting, resumption and the evaluation gate possible. If you need arbitrary code, Modal is the honest recommendation.
Why is your GPU-hour rate lower?
Because we source capacity as serverless compute across multiple providers rather than from one, and we schedule against a class rather than a specific machine. What that buys is a lower rate for the same GPU class; what it costs is that you choose a class, not a host.
Could I use Modal for training and Onrup for serving?
You could, but there is little reason to. If the model was trained elsewhere, its adapter was not produced by a template we validated, and most of what our serving layer adds — the gate, the cost meter, the lineage — depends on that.
Researching Modal alternatives more broadly? →
Last verified 6 August 2026. Modal figures come from their own pricing page on the date checked.
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.