Training

Training overview

Adapt a base model to your task by training it on your own data, so it learns your formats, tone, tools, and edge cases instead of relying on prompt instructions alone. Reka runs the training for you, without the burden of building and maintaining your own GPU or training infrastructure.

Note Training is run with the Reka team rather than self-serve. If you have a task where a general model falls short, or a prompt that has grown too long to be economical, get in touch and we’ll scope a fine-tune with you.

When training is worth it

Training is worth it when you want:

  • Higher task quality — Beat a general-purpose model on your specific workload, and often match or exceed a larger closed model.
  • Lower latency and cost — A smaller specialized model can replace a bigger one at a fraction of the per-token cost.
  • Consistent behavior — Bake in formats, style, and tool use so you stop paying for long prompts and few-shot examples on every request.
  • Ownership and no infra — You keep the resulting weights, and Reka handles the GPUs, scheduling, and checkpointing.

How it works

  1. Tell us about the taskThe base model you have in mind, what the model should do, and what data you have: rough size, format, and whether it already includes the outputs you want.
  2. We train and evaluate itThe Reka team prepares the data with you, runs training on our infrastructure, and shares evaluations against your baseline before anything ships.
  3. Serve it like any other modelThe fine-tuned model is deployed on dedicated capacity and called through the same OpenAI-compatible API, so your existing code keeps working.

Before you train

Fine-tuning pays off once prompting has been pushed as far as it goes. A few things that are usually faster to try first:

If those don’t get you there, that’s a good sign training will. and we’ll help you decide.

Next steps