Skip to main content
Prompt Baking encodes prompt behavior directly into model weights. Fine-tune AI models with the convenience of prompt engineering. You shouldn’t need a PhD to teach your AI what it needs to know or wait for someone else’s model update to give you what you need. With Bread, your model permanently inherits the prompted behavior with zero input tokens at inference time.
The value: After baking, your model can exhibit prompted behavior with zero input tokens. The skills live in the weights, not in runtime prompts.

How does Baking work?

Learn the four-phase process to baking

What is bgit?

Learn the git-native interface to baking

Quickstart

Run a full bake in 5 minutes

Advanced Features

Configure generators, targets, and advanced baking options

​
Core Concepts

Repositories

Like GitHub, these repos hold all of a model’s bake data

Prompts

The actual prompts you want to “bake in”

Stim Data

User prompts the prompted model would receive

Rollout Jobs

Responses from the prompted model to the stim data

Targets

The composition of stim & rollout for a given prompt

Bakes

Configuration for your bake with rollout data

​
Full Process: Prompt, Stim, Rollout, Bake

1

Prompt

Specify the prompts to bake in
2

Stim

Run stim jobs to create synthetic dataset
3

Rollout

Run rollout jobs to get model responses to stim
4

Bake

Configure and run bake jobs for model training

​
Requirements

  • Python: 3.8 or higher
  • Dependencies: httpx (included), optional aiohttp for async performance
  • API Key: Required for authentication

​
Need Help?

GitHub Repository

Check out our public repo to start baking today

Workflows Guide

Learn common patterns and workflows

Error Handling

Handle errors gracefully

Client Configuration

Unlock powerful SDK capabilities