AnswersCustom models
In short
Train your own model for the narrow, high-volume work that defines the business, where a smaller model trained on your data can match the frontier at a fraction of the cost, or where regulation keeps frontier models out. For everything else, route across frontier and open-weight models by task. Most companies start with applications and agents, and train once their data and evaluations build up.
| Situation | Why a model of your own helps |
|---|---|
| High volume, narrow task | A small trained model matches a frontier model on that task at a fraction of the cost and latency. |
| Proprietary judgment | The model learns decisions only your experts make, from your own records. |
| Regulation or residency | A model you run in your environment satisfies rules a hosted frontier model cannot. |
| Access you control | No supplier can change its terms, retire the model or stop serving you. |
| The hardest, broadest work | Keep it on the best frontier models, routed per task. |
AI application companies train models to stop depending on labs that now compete with them. Enterprises do it to keep an edge no competitor can rent. The companies getting results own a few narrow, high-volume tasks and still use frontier models for the hardest work.
Everyone wants their own model →Should I train my own model? →
An evaluation set built from real cases, enough examples of the work done well, and a clear task. Those come from running the work in production first, which is why training usually follows the application layer.
Fig partners at every layer: applications and agents on frontier models first, then the company's rules and evaluations codified into CosmOS, then models trained on its data with supervised fine-tuning, preference tuning, reinforcement learning and distillation, chosen per task. The company keeps the weights.
It depends on the task. Narrow tasks can improve with thousands of good examples, and an evaluation set of real cases is as important as the training data.
Sometimes. Preference tuning, reinforcement learning and distillation each push quality or cost further, and the right mix depends on the task.
Retrain on the newest open base models as they ship, and keep routing the hardest work to frontier models.
Dream big, build fast, and grow far with Fig