Introducing Fine Tune: replace an LLM call with a small model you own

Lots of apps ask a large language model the same small question thousands of times a day. Is this message spam? Which team gets this ticket? Is this review positive? The answer is one word from a short list, but every call still pays for a big model, waits for it, and sends the text to someone else's servers.

Fine Tune turns that call into a small model you own. Upload a file of your text, say what to decide, and minutes later you have a model that answers the same question on a CPU. Run it on our servers for a fraction of a cent, or on your own machine for free.

One spam filter, start to finish

The video follows one real job, from a Word file to a model that needs no LLM at all.

1. The file

   A Word file with 2,300 text messages and no labels.

2. The question

   Spam or not spam.

3. The labels

   An LLM labels every message once. After training, the filter never calls an LLM again.

4. The model

   Those labels train a private copy of OpenDecision-Small.

5. The test

   240 messages are held out of training and used to score it.

It calls the prize text spam at 99.9999%. That 99.2 measures how closely the small model copies the LLM's answers, which is the point: it makes the call your LLM made, without the LLM. For a score on ground truth, answer some rows yourself.

What changes when you own the model

On our servers a call bills at least 256 tokens, so a million short messages sent one at a time cost $6.40. On your machine, decideLocal runs the same model file for free and the text never leaves it. Only your organization can call the hosted copy.

Try it on your own file

Any LLM call that picks from a short list is a candidate: spam, sentiment, routing tickets to a few queues, flagging risky requests.

1. Open Fine Tune

   In the dashboard, choose New fine tune.

2. Say what to decide

   Name the model and write one question, like Which team should handle this ticket?

3. Add your file

   CSV, TSV, JSON, JSONL, Excel, Parquet and plain text work on every account. After a paid top-up, an agent converts other formats, like the Word file in the video.

4. Answer 20 rows you know

   For a CSV, TSV, JSON, JSONL or text file, the form shows its rows one at a time. These answers help more than anything else you can add.

Jobs took 3 to 10 minutes in our tests, and we email you when yours finishes. Then a call needs only the model's name and the text:

import { Tokz } from "@tokz/sdk";

const tokz = new Tokz({ apiKey: process.env.TOKZ_API_KEY! });
const text = "I was charged twice.";

const hosted = await tokz.decide("support-triage", text);
const local = await tokz.decideLocal("support-triage", text);

Python has the same two calls.

How much does Fine Tune cost?

All our prices are on the pricing page.

When does fine tuning help?

Tuning wins most with a few labels and real text, where the stock model is weak. On public datasets:

Macro F1 out of 100 on 200 held-out rows, one run each, measured on 2026-11-03. Reruns move these by a few points. We ran 10 datasets. On the other 6, mostly many labels, long text or a messy file, five showed no clear difference and one came out worse. The docs have all 10. That is why a fine tune that does not win is free.

Your data stays yours

Your text goes to RunInfra, Freestyle and Modal to build the model, and it trains only your model. We hold your files in memory until the training machine has them, so they never reach our disk, our database or our logs.

Have an LLM call that answers from a short list? Start with $5 of free credit and upload the file. If the model does not win, the training costs you nothing.