Fine-tuning

Teach an open model the job you do every day

A big cloud model is good at everything and expensive at all of it. Give Geyser a few hundred examples of one job done right, and it fine-tunes a small open model to do that job your way. The model runs on your own computer, so every answer after that is free and private. Start from open models on Hugging Face, and train on your own Mac or PC, or on Tinker from Thinking Machines.

“June answered four hundred refund emails last quarter. We taught a small model from them, and now I sort the new ones with it on the office Mac.”

Pia
How it works

From a pile of examples to a model your Agents use

You can do it yourself with “Teach a new model”, or ask an Agent to run the whole project. The Agent gathers and checks the examples, writes the test, trains, and comes back with the scores for you to accept.

  1. 1
    Describe one job

    Say what goes in and what should come out: sort a support email into one of your eight queues, or pull the invoice number, total, and due date out of a PDF. A narrow job trains well on a small model.

  2. 2
    Give it examples you’ve checked

    Type them, paste them, or upload a CSV or JSONL file, from 5 examples up to 2,000. Geyser leaves out examples that look like they hold personal details, removes repeats, and keeps a fifth of them back for the test.

  3. 3
    Train it

    Training runs on one of your own computers at no charge, or on Tinker from Thinking Machines when the job needs a bigger model. Either way it’s LoRA fine-tuning, so the result is small and quick to load.

  4. 4
    Test it against where it started

    The new model answers the examples it never saw, and so does the model it started from. You see both scores. If the new one doesn’t do at least as well, nothing changes for your Agents.

  5. 5
    Put it to work

    Make it the main model for the Agents you choose, let your Agents ask it whenever that job comes up, or call it from your own app.

Where it trains

Your graphics card, or Tinker when you need more

Most narrow jobs train well on the computer you already have. When one needs a bigger model or a different kind of training, Tinker does the heavy lifting and Geyser keeps the examples, the test, and the bill in order.

On your own computer

A Mac with Apple silicon trains with Apple’s MLX. A Windows or Linux PC with an NVIDIA graphics card, or a Linux PC with Intel graphics, trains with Hugging Face’s Transformers and PEFT libraries. Training waits until the computer is plugged in and cool, and you can keep it to the hours you choose.

  • No charge, on every plan
  • Your examples go to no outside company
  • Starts from Qwen 3.5 on a Mac and Gemma 4 on a PC

On Tinker, from Thinking Machines

Tinker trains bigger open models without you renting a single GPU, and it goes beyond examples and answers. An Agent can use it to teach a model from your preferences between two answers, from a bigger model’s answers, or with reinforcement learning.

  • Qwen 3.5, Qwen 3.6 35B A3B, and gpt-oss 20B
  • You see an estimate and set a spending cap before it starts
  • Tinker can host the model, or you bring it home to a Mac
Who we build on

Standing on Hugging Face and Thinking Machines

We’d rather build on the best tools in open AI than make our own copies of them. These two companies do a great deal of the hard work behind every model you teach in Geyser.

Hugging Face

Every model you train on your own computers starts from the Hugging Face Hub, pinned to an exact version so a retrain starts from the same place. PCs train with Hugging Face’s Transformers and PEFT libraries. And to run a model without training it, you can search the Hub from inside Geyser and see which of your computers it fits.

Thinking Machines

Tinker, from Thinking Machines, lets an Agent fine-tune models far bigger than your computer could hold, and handles the GPUs, the scheduling, and the failures for you. Connect your own Tinker account once, and Geyser trains in a project of its own for your workspace.

Your examples stay yours

What happens to the examples you give it

Your examples are your business’s know-how, so Geyser treats them that way. You decide where training happens, and the model that comes out belongs to your workspace.

How privacy works in Geyser
  • Training on your own computers sends your examples to no outside company.
  • Tinker sees your examples only when you allow it as an outside provider, and data you’ve marked for your own computers never goes to it.
  • Mail and files from Google Workspace and Microsoft 365 aren’t used as training examples.
  • Your examples train your models and nobody else’s. They’re never pooled with another customer’s.
  • Delete the examples a model learned from, and that model stops being used.
Use it anywhere

Call your model from your own app

Open the model to one of your developer projects and call it with the OpenAI or Anthropic SDK you already use. Requests reach it through Geyser, so your app works from anywhere without opening a port on your network, and a model on your own computer still costs nothing per request.

Build apps on a model you taught
Python · OpenAI SDK
import os
from openai import OpenAI

# GEYSER_API_URL is the API URL shown when you create the key.
client = OpenAI(
    base_url=os.environ["GEYSER_API_URL"] + "/api/v1/openai",
    api_key=os.environ["GEYSER_API_KEY"],
)
reply = client.chat.completions.create(
    model="private/pmd_…",
    messages=[{"role": "user", "content": "Summarize this ticket"}],
)
Good questions

Questions about fine-tuning

How do I start?
Open Models in Geyser and choose “Teach a new model” to do it yourself, or “Build a model” to have an Agent run the project. To train on Tinker, first connect your Tinker account in Settings, under Accounts.
What does it cost?
Training on your own computers is free, and so is every answer the model gives there. Training on Tinker runs on your own Tinker account, so Thinking Machines bills you directly. You see an estimate and set a spending cap before it starts.
How many examples do I need?
Five is enough to try it, and a few hundred good ones is a better start. The more consistent your examples are, the better the model gets at the job.
Is it full fine-tuning?
No. Geyser trains a LoRA adapter on top of an open model, which needs far less memory and time. On Tinker an Agent can also use preference training, distillation from a bigger model, or reinforcement learning.
What if the new model is worse?
Then nothing changes. A model only goes to work when it does at least as well as the one it started from on examples it never trained on, and you can roll back to an earlier version.
Can I run a model from Hugging Face without training it?
Yes. Search the Hugging Face Hub from Geyser, pick a model, and Geyser tells you which of your computers it fits before anything downloads.
Make a start

Teach a model one job this week

Pick the task your team repeats most, gather a few hundred examples you trust, and see how close a small model on your own computer gets.