Distributed AI training company Prime Intellect is taking its agent post-training platform Lab out of beta and into general availability. Lab bundles evaluation, reinforcement learning (RL) training, adapter deployment, and inference into a closed loop: you define a task and scoring criteria, and the platform automatically drives the model through trial and error, collects reward signals, trains LoRA adapters (a parameter-efficient fine-tuning method that updates only a small fraction of weights), and redeploys the new weights for another iteration.
The platform's core abstraction is an Environment, which packages task data, a model harness, a sandbox, and reward metrics. The same Environment can be reused for local development, hosted evaluation, synthetic data generation, and RL training. Training is charged per token, not per GPU hour, and is powered by the company's open-source prime-rl framework. At launch, the GA supports 14 models from Nvidia, OpenAI, Meta, and Qwen, ranging from 1B to 70B parameters, covering both dense and MoE architectures.
Prime Intellect was founded in 2023 by Vincent Weisser and Johannes Hagemann and has raised over $70 million to date, with a Series A led by Founders Fund and a Series B led by Radical Ventures. The company is best known for its open-source distributed training work on frontier models, including INTELLECT-3, a 106-billion-parameter MoE model. During the beta, hundreds of users ran over 10,000 training tasks on the platform.