Fine-tuning
A model that knows your work.
Teach an open model your organisation's language, documents and rules, then serve it from the same place. Your data and the resulting model stay yours, in Nepal.
Who it's for
Models that need your terms, tone or language.
Banks and finance
A model that understands your products, forms and compliance language.
Public sector
Models tuned to Nepali documents and services, without data leaving the country.
Product teams
A smaller, cheaper model that does one job as well as a big one.
What teams use it for
- An assistant that answers from your policy manuals
- A classifier for complaints, by product and urgency
- A Nepali-language model for your own domain
How it works
Three steps.
- 01
Bring your data
Examples of the work, uploaded once to storage beside the GPUs.
- 02
Tune an open model
In a notebook or a pipeline, on whole GPUs or slices.
- 03
Serve what you made
Put the tuned model behind an endpoint, on the same platform.
What you get
Fine-tuning, in detail.
Start from an open model
Begin with a capable open model rather than training from scratch.
Notebooks or pipelines
Experiment in JupyterLab, then repeat the run as a Kubeflow pipeline.
Every run tracked
Settings, data versions and results are recorded in MLflow, so you can compare and repeat.
Your weights are yours
The data you bring and the model you make belong to you. We never use them for anything else.
Serve from the same place
No export and re-upload: deploy the tuned model with managed inference.
Pay by the second
Fine-tuning runs on on-demand GPUs, billed per second in rupees.
Under the hood
The key numbers.
- GPU
- NVIDIA H200 NVL
- GPU memory
- 141 GB
- Billing
- Per second
- GPU
- NVIDIA H200 NVL
- GPU memory
- 141 GB
- Billing
- Per second
- Memory bandwidth
- 4.8 TB/s
- Currency
- Nepalese rupees (NPR)
- Ways to pay
- Digital payment, bank transfer or invoice
Billing
GPU time used, per second, in NPR
Commitment: none. Rates are being finalised; every request gets a firm quote in rupees.
How billing worksQuestions
About fine-tuning.
How much data do we need?
Often less than people expect: hundreds to a few thousand good examples. We can help you judge yours.
Who owns the fine-tuned model?
You do. Your data and the resulting weights are yours.
Can you do it for us?
Yes. Our engineering services can plan and run the fine-tune with your team.
Start with fine-tuning.
Tell us what you want to run. It takes about a minute.