An on-demand lab to develop and improve enterprise AI models

Training, deployment, and production feedback come together in Model Factory, Prime Intellect’s new platform.

Training, deployment, and production feedback come together in Model Factory, Prime Intellect’s new platform.

A failed tool call, an ineffective patch, or an abandoned task can become material for another training run. This connection between real-world usage and weight updates is central to Model Factory, Prime Intellect’s platform bringing together model training, evaluation, and deployment.

The offering extends the company’s previous Lab service, expanding the work teams can perform on models, from pretraining to supervised fine-tuning and reinforcement learning. Teams define their tasks, data, and success criteria; the platform orchestrates compute, inference, and execution environments. A trained model can then be deployed as an application endpoint, used to generate synthetic data, or submitted for another evaluation. Different ways to update model weights

Supervised fine-tuning adapts a model using selected examples: expected responses, demonstrations, or successfully completed tasks from a particular domain. It can serve as a first step before reinforcement learning. In that second stage, an agent attempts tasks, receives a score, and has its weights updated based on those attempts. The team running the experiment chooses the reward signal.

For agents that write code or use tools, execution takes place in isolated environments. Integrations with verifiers, the Environments Hub, and Harbor connect tasks to the available tools and scoring methods. The announcement describes support for up to tens of thousands of concurrent microVM sandboxes, a scale claimed by the provider.

Pretraining and continued training of existing base models round out these capabilities. The documentation nevertheless distinguishes the platform’s broader ambitions from the configurations already supported: hosted pretraining remains in beta, with training from randomly initialized weights tested on dense architectures. Hosted supervised fine-tuning is also labeled beta. From managed training to a dedicated cluster

Two access options provide different levels of control. With the on-demand service, a team submits its configuration and uses GPUs managed by Prime Intellect. Billing depends on the GPUs occupied and the time they are used. Persistent storage and trace ingestion incur separate charges.

Dedicated deployments are set up with the provider’s team. They bring training, inference, Slurm, and Kubernetes together on the same cluster. Direct infrastructure access supports deeper changes, such as selecting a specific version of the training software rather than using the default managed build.

The promise of control also extends to the resulting models. The service terms assign ownership of trained models to the customer, subject to the rights attached to the base model and third-party components. Starting with an open weights model still requires compliance with its license. Collecting traces does not guarantee improvement

Recording task attempts connects the deployed model to the next experiment. Traces capture the tools called, the path taken, and the points of failure. Reviewing them helps teams select new examples, build an environment, or revise how a task is scored. That selection remains a research and engineering responsibility.

For a coding agent, a rising score can be checked against the patches it produces and the results of its tests. Intermediate model checkpoints can then be evaluated on tasks separate from those used for training. The team can use that evidence to decide whether the new version should replace the one running in production.