PRIME-RL
#11 v Frameworky pro agentyPrime Intellect · 2× · naposledy 08. 8. 2026
PRIME-RL is an Open Source framework from Prime Intellect for large-scale, agentic Reinforcement Learning of language models. It combines training, Inference, and orchestration capabilities, enabling asynchronous, distributed RL training—from single-GPU setups to multi-node clusters via Slurm/Kubernetes. Starting with version 0.8.0 (alongside the "verifiers" library 0.3.0), PRIME-RL supports multi-agent systems for the first time, where multiple agents collaborate or compete in programmable interactions with credit assignment across the entire interaction. The framework integrates natively with the Environments Hub platform and is used, among other applications, for training models such as INTELLECT-3.