The Code/X ArchiveView on X
Prime Intellect

@PrimeIntellect

Introducing Model Factory.

Train, deploy, and continuously improve your own models on one stack.

Run on-demand, on dedicated clusters, or bring your own.

Own your intelligence. Own your model factory.
49104925520
Prime Intellect

@PrimeIntellect

Frontier labs own the full loop, training and deploying intelligence, and capturing the data required to train the next model.

We believe every company should have that same ability.

Model Factory brings that loop into one open, customizable stack. Here's how it works:
1712638
Prime Intellect

@PrimeIntellect

Train agents with reinforcement learning.

+ Large-scale async RL with managed prime-rl
+ Your tasks, environments, and rewards
+ verifiers, Environments Hub, and Harbor support
+ Thousands of concurrent sandboxes for agents that write code and use tools
11546
Prime Intellect

@PrimeIntellect

Pre-train. Mid-train. Run your own research.

+ Scale to thousands of GPUs with automatic recovery from node failures
+ Train efficiently with Prime Mega MoE: 48% MFU on H200s with Qwen’s 235B MoE
+ Explore new architectures and objectives, run ablations and launch autonomous training experiments
11544
Prime Intellect

@PrimeIntellect

Fine-tune on your own data.

+ Run supervised fine-tuning with hosted SFT
+ Evaluate checkpoints throughout training to see whether your model is getting better at your tasks

Build the starting point for your next RL run.
11464
Prime Intellect

@PrimeIntellect

Agents and RL need somewhere safe to run.

+ Model Factory integrates Prime Sandboxes: hardware-isolated microVMs you can spin up, checkpoint, fork, and lock down
+ Built for RL, they easily scale into the tens of thousands
11412
Prime Intellect

@PrimeIntellect

Deploy the model you trained.

+ One click from checkpoint to inference endpoint
+ Run evaluations and synthetic data generation at scale
+ Training and inference on the same infrastructure
11403
Prime Intellect

@PrimeIntellect

Turn production experience into your next training run.

Enable trace collection with Prime Traces. Inspect what your agents actually do, find where they fail, and turn those failures into environments.

Deploy → collect traces → evaluate → train → repeat.
11422
Prime Intellect

@PrimeIntellect

The foundation: self-healing compute.

+ Health checks and automatic node replacement
+ Automatic recovery from the latest checkpoint
+ Move GPUs between workloads through the prime factory CLI

Dedicated Model Factory deployments bring training, inference, Slurm and Kubernetes access onto the same fabric, across all your compute.
11443
Prime Intellect

@PrimeIntellect

We’re already deploying Model Factory for teams working on finance automation, computer use, scientific discovery, chip design and more.

Their data. Their environments. Their models. A factory built around what they need.

Starting today, we’re putting those same capabilities in your hands.
Image from the post
11402
End of thread