Docs

Fern models

The three Fern models, what each is suited to, and how to choose one for a project.

Every project fine-tunes one of three Fern models on your tools. They differ in size: a larger model is more accurate, and slower to train and to answer.

ModelSizeModel idSuited to
Fern Bud12M parametersfern-bud-1.0High volume and small tool sets. Trains in minutes.
Fern Pinnule160M parametersfern-160mTool sets that have grown past what Fern Bud handles well.
Fern Pinna1.5B parametersfern-1500The highest accuracy. Used by the intent bar start and the templates.

A parameter is one of the numbers a model learns during training. The count is a rough measure of how much a model can hold: Fern Pinna has about 125 times as many as Fern Bud.

Which model a project uses

  • From scratch starts on Fern Bud unless you choose another.
  • Intent bar for your app always trains on Fern Pinna.
  • Templates train on Fern Pinna, with settings tuned for the template.

To choose, open Advanced mode on the wizard's Train step and pick a card under Base model.

A project's base model is fixed by its first training run. Every later run, including retraining, uses the same model and the same settings. To try a different model on the same tools, create a new project.

Choosing a size

Start with Fern Bud if your app has a few dozen tools at most and you want quick training runs while you iterate. Move to Fern Pinnule or Fern Pinna when:

  • the model often picks the wrong tool between two that do similar things,
  • it misses arguments in longer sentences, or
  • your users write full, varied sentences rather than short commands.

The playground runs the same requests through Fern models and GPT-4o side by side, so you can see how the sizes compare on a calendar and a CRM tool set before you train.

Where the models run

Fernfly hosts every deployed model and runs it on GPUs. You call it through your project's endpoint, so there is nothing to install or scale.

Fern Bud is small enough to run on ordinary CPU hardware, which makes private and offline deployment possible. The Download model button on a project's Overview doesn't download a file yet: it asks you what you would use an offline model for and sends us the request. For a private deployment today, contact us.

Revised

Was this page helpful?