Fine-tuning in a world of APIs

When getting a trained model is an API call away, why do we need to fine-tune ourselves anymore?

It’s 2025, every single provider out there offers ways to fine-tune your closed (or open source) model on your data through a paid API schema. These typically are in the ballpark of either a per token cost (OpenAI is ~$0.008-0.012/1K tokens) or an hourly cost (Fireworks is ~$0.50-5.00/GPU hour).

You’re paying an increase in price to have the convience of not needing to know what’s going on during training and simply pointing your data at an endpoint, and a few hundred (or thousands) of dollars later you have a custom model.

This is great, right?

Right?

What you sacrifice

Convience needs to sacrifice something to occur. Usually this is in the form of control. You don’t know how the sausage gets made. You’re assuming that these fancy labs do. And while that’s good for some people, if I’m spending a hefty chunk of money to train my model in a day-and-age when open source alternatives to closed source labs are growing closer and closer to being equivalent, there is little reason for charging more for me to make use of data and tools I could use instead myself.

What if their training has leakeage? You don’t know. Nor do you know if they’re doing anything with your training data.

When it comes to situations like these, getting your hands dirty to get that training done is more often better than not.

What’s the best way to get started?

Let’s assume you want the minimal headache possible to get the ball rolling. You’ve got your data, a model, and you just want the training done. But you want it done on your own (rented) hardware, and appraocable enough that you can figure out what’s going on.

My personal tier list of frameworks I’ve used is either going to axolotl or torchtune (Will agrees), however Unsloth is also a great direction for starting out since their entire framework is around making model training as accessable as possible through Google Colab (a freemium GPU provider).

In a world when open-source models are on the state-of-the-art, learn how to work them to your advantage and get the job done, while also being able to maintain complete control of your model from start to finish.