SunderSystems

Home / Distillation

AI model distillation for one job

We take a leading open-weight model and train a much smaller one to do a single task in your company as well as the large one does. The aim is a model that runs on a modest server and does not send prompts to an outside provider. Per-client distillation is in development, and we have not yet delivered a client model.

What it is

Distillation trains a small model, the student, to reproduce how a larger model, the teacher, behaves on a specific job. We use open-weight teachers such as Qwen, Llama, DeepSeek and Mistral, and we check the licence of each one before it goes near your project.

The student is narrow on purpose. It does one thing, such as classifying support tickets, extracting fields from contracts or drafting replies in your house style, and it does that one thing quickly on small hardware.

Who it is for

Companies that already use a hosted AI service for a repeated, well-defined task and want to stop sending that work, and the data in it, outside the building. It suits teams with a few hundred to a few thousand examples of the task done well, or who can produce them with our help.

What you would get

  • A trained model (planned). Handing over the weights is in development. When it launches, they will come with the licence terms of the base model written out in plain language.
  • An evaluation report. Scores for the student and the teacher on a test set built from your own examples, with the failures listed, not hidden.
  • The training recipe. The data preparation, training configuration and test harness, so your team or another vendor can reproduce and extend it.
  • A deployment handoff. Everything needed for the installation or hosting step, which we can also do for you.

How it is done

  1. Define the task. We write down the inputs, the outputs and what counts as a correct answer.
  2. Build the test set. We set aside examples the model never trains on. This is the yardstick for everything after.
  3. Train the student. The teacher labels and explains your examples, and the student learns from them. We iterate on data quality before model size.
  4. Measure and decide. You see the scores next to the teacher's. If the student falls short, we say so and adjust the scope before you spend more.

What we need from you

  • One task with a clear definition of a good answer.
  • Example inputs and, ideally, good outputs. We will work with what you have.
  • A person who can judge the model's answers during testing.
  • Agreement on where the data may be processed. Work on your premises can be scoped. We have not yet delivered it.

How we handle your data

  • Only with your consent. We train only on data you have consented to. A workspace owner turns training on, and we record that choice.
  • Open-weight teachers only. Text written by Claude or any other closed model is never used for training.
  • Imported chats stay out. Chat histories you import from other tools are stored read-only and are never used for training.
  • Your model is yours alone. Training runs on our own hardware in the Philippines. We test each model against a gate before it goes live, and it is served only to your workspace. Your data is never used to train a model for another client.
  • You can delete it. “Delete model and data” in the app removes your models, examples and working notes. See our privacy policy.

Last updated: 17 November 2026

Questions, answered.

Is a distilled model as accurate as the large one?

On a narrow task it can match it, but we do not promise that in advance. We test both on your held-out examples and show you the scores before anything is installed.

How much data do we need?

It depends on the task. Often a few hundred to a few thousand good examples are enough to start. When you request training we review what you have and tell you what is missing.

Other services

  • InstallationOn your own servers. Scoped per client, not yet delivered.
  • HostingA hosted AI desktop on hardware we operate.
  • Agentic systemsAgents for multi-step work, under your approval rules.

Start on a practice company.

Request an invite and try it on made-up customers first. Nothing reaches real people while you learn.