Prime Intellect launched Model Factory on November 18, pitching it as the same build-deploy-learn cycle that frontier labs run internally, now open to any company that wants its own models. The platform covers reinforcement learning (RL), supervised fine-tuning (SFT), pre-training, and inference from a single control plane. Early deployments, according to the company, span four fields: finance automation, computer use, science research, and chip design.

The argument behind the product is about ownership. A company that rents a model through an API cannot train it against its own reward signal, because it never holds the weights. Prime Intellect also says the provider sees the prompts and data, which lets it learn the customer’s domain, and that the vendor may pull or swap the model version a product relies on, giving customers little notice. Owning an open-weights model removes those three exposures, but the company then has to run the machinery.

That machinery is the real barrier. Prime Intellect describes it as sourcing GPUs, writing a custom training framework, keeping inference running at scale, and operating large sandbox fleets. Model Factory packages those pieces. It succeeds Lab, the company’s earlier platform, which Prime Intellect says could not take a model from pre-training all the way through large-scale RL.

The training loop it automates is simple to state. An agent attempts tasks and gets scored. An orchestrator batches those scored attempts and hands them to a trainer that updates the weights. The updated model tries again. Each cycle needs inference for responses, sandboxes where code and tool calls execute, and environments that grade the output. Researchers specify tasks, rewards, and the training recipe. The platform takes over the trainer, the generation of attempts, and the coordination between them.

Several integrations are named. Model Factory works with verifiers, the Environments Hub, and Harbor, and a run can attach microVM sandboxes by the tens of thousands, all active at once, when the agent writes code or calls tools. Teams choose a model from the open-source catalogue or supply one. For pre-training, Prime Intellect cites its Prime Mega MoE reaching 48% model FLOPs utilization (MFU) on H200s with Qwen’s 235B mixture-of-experts architecture. That figure comes from the company’s own tests; no independent benchmark accompanies the announcement.

The loop closes in production. Deployment is a one-click inference endpoint from the dashboard, and training and serving share infrastructure, so the same model can power an application, generate synthetic data, or serve as the baseline for the next evaluation. A feature called Prime Traces logs every tool an agent invoked, the route it followed through a task, and the point where it broke. Those failures feed the next environment and the next round of training.

Reliability gets its own section for the dedicated tier. Long experiments hit hardware faults, so the dedicated tier monitors node health, swaps out failing nodes automatically, and resumes from the most recent checkpoint. A prime factory command-line tool gives teams one interface for managing compute and shifting GPUs between workloads.

There are two ways in. On-demand hosted training takes a configuration and executes it on hardware Prime Intellect manages, with charges counted in GPU-hours. Dedicated deployments bring reserved capacity, managed training and inference, and direct Slurm and Kubernetes access for teams that need to change frameworks, architectures, objectives, or rollout strategies. Prime Intellect’s blog post does not list prices for either option.

The pitch rests on a claim the announcement does not test: that a mid-sized company, with its own tasks and rewards, can beat a general-purpose provider in its niche. The post offers no customer results, no cost comparison against renting, and no data on how much production traces improve a second model version. Those are the numbers a finance or chip-design team would need before moving a workload off an API.

For teams already paying a lab to fine-tune a closed model, the practical step in the next ninety days is a small pilot: pick one task with a measurable reward, run it on the on-demand tier, and compare the result against the incumbent API before asking for reserved capacity.

Prime Intellect, “Introducing Model Factory,” company blog, November 18, 2026.