PRIME-RL Launches Open-Source Framework Scaling Reinforcement Learning to 1000+ GPUs and Trillion-Parameter Models

PRIME-RL launches as a fully open-source reinforcement learning framework capable of scaling to 1000+ GPUs and trillion-parameter models, supporting major AI model families with asynchronous training, multimodal capabilities, and flexible multi-node deployment under the Apache 2.0 license.

• PRIME-RL is an open-source framework for large-scale reinforcement learning, designed to scale to 1000+ GPUs with fully asynchronous RL training, supporting models over 1 trillion parameters using FSDP2 and vLLM with FP8 inference and multiple parallelism strategies.
• The framework supports a wide range of model families including Qwen3 MoE, GLM-5, MiniMax M2, Nemotron H, and others, with native integration for agentic and SWE environments, multimodal VLM support, and end-to-end post-training capabilities including SFT, RL, and evaluations.
• PRIME-RL offers multi-node deployment via Slurm and Kubernetes, hands-on training examples ranging from single-GPU toy tasks to 2048-GPU production runs, and is fully open-source under the Apache 2.0 license with active community contributions and 1.4k GitHub stars.