Prime Intellect Releases prime-rl 0.6.0 to Train Trillion-Parameter MoE Models on Agentic RL Workloads

Prime Intellect has unveiled prime-rl 0.6.0, an open-source framework designed to enable asynchronous reinforcement learning (RL) training on trillion-parameter Mixture-of-Experts (MoE) models. The release represents a significant advancement in scaling agentic AI systems, demonstrating unprecedented capabilities for training large language models on complex, real-world tasks at massive scale.

Prime Intellect's prime-rl 0.6.0 successfully trained the GLM-5 model on software engineering (SWE) tasks, achieving remarkable performance metrics across distributed infrastructure. The framework processed sequences up to 131,000 tokens in length while maintaining sub-5-minute step times and executing 256 rollouts simultaneously on a cluster of 28 H200 nodes. These specifications indicate substantial progress in handling the computational complexity traditionally associated with training trillion-parameter models using RL algorithms.

The framework's asynchronous architecture enables efficient resource utilization across distributed environments, addressing one of the primary bottlenecks in large-scale model training. By supporting MoE architectures—which activate different model components for different inputs—prime-rl 0.6.0 improves training efficiency and scalability compared to dense model alternatives.

• Democratized Large-Scale Training: Open-sourcing the framework enables organizations beyond major AI labs to experiment with trillion-parameter model training
• Agentic AI Development: Enhanced capability for training models on agent-based tasks, essential for autonomous AI systems and complex problem-solving applications
• Infrastructure Optimization: Sub-5-minute step times demonstrate practical feasibility of RL training on enterprise-scale hardware clusters
• Competitive Acceleration: Removes technical barriers that previously limited RL exploration at the trillion-parameter scale

The release of prime-rl 0.6.0 addresses a critical gap in the AI development landscape. Training trillion-parameter models using reinforcement learning has remained largely inaccessible due to computational complexity and proprietary tooling constraints. By providing an open framework that delivers proven results on demanding SWE tasks, Prime Intellect enables broader innovation in agentic AI systems—models capable of autonomous reasoning and task execution.

This advancement carries implications for AI safety research, enterprise AI deployment, and the competitive dynamics of AI model development. Organizations can now more readily experiment with advanced training methodologies previously restricted to well-resourced institutions, potentially accelerating progress toward more capable and reliable AI agents.