Model Factory Overview

Train, evaluate, deploy and improve your own models on one platform

Model Factory is where you build your own models on Prime Intellect. You train a model, evaluate it, deploy it, and use what it does in production to train the next version. Each step runs on the same platform, so a checkpoint from one step is ready for the next.

Model Factory is built on our open-source libraries: prime-rl for training and verifiers for environments and evaluations.

​

The loop

Step	What you do	Where
Train	Run reinforcement learning, supervised fine-tuning or pretraining on an open-weights model.	Hosted Training
Evaluate	Measure a model on environments from the Environments Hub, or on your own.	Hosted Evaluations
Deploy	Serve a model as a dedicated inference endpoint.	Dedicated Endpoints
Improve	Capture traces from your runs and endpoints, find where the model fails, and feed that into the next run.	Prime Traces

​

Two ways to run it

	On-demand	Dedicated cluster
What it is	You submit a config and we run it on GPUs we operate.	A cluster reserved for your workloads alone.
Access	Self-serve	Contact us
Billing	Per GPU-hour, prorated to the second	Reserved capacity
Use it when	You want to start a run without reserving anything.	You need control over the stack: Slurm and Kubernetes access, or a pinned build of the training framework.
Start here	Getting Started	Model Factory Clusters

To run the same training framework on your own hardware, use prime-rl directly.

​

Train

Hosted Training supports three kinds of run. Each one is dispatched with prime train and a TOML config.

Reinforcement learning

Full-parameter RL against a verifiers environment.

Supervised fine-tuning

Train on examples of the behavior you want.

Pretraining

Pretrain a model from scratch on your own corpus.

Run prime train models to see which models start warm and what each GPU type costs. See Models for details, and Volumes for how checkpoints are stored.

​

Evaluate

Run an environment against a model from the dashboard or with prime eval run --hosted. Results appear on the Evaluations page. See Evaluating environments for the local and hosted workflows.

​

Deploy

Create a dedicated inference endpoint from the Dedicated Endpoints page. An endpoint serves a model through an OpenAI-compatible API on GPUs reserved for it. For shared, pay-per-token models, see Prime Inference.

​

Improve

Prime Traces stores the traces from your training runs, evaluations and endpoints. Use them to see what the model did on each task and where it failed. A recurring failure can become an evaluation or a training task in your next run.

Trace uploads incur ingestion and storage charges. See Pricing and settings.

​

Next steps

Getting Started

Launch your first on-demand training run.

End-to-End Run

Full walkthrough with baseline evals, model selection and result review.

Model Factory Clusters

Dedicated clusters running training, inference, Slurm and Kubernetes.

Platform Overview

How environments, training, evaluation and inference fit together.