prime-rl

prime-rl is Prime Intellect’s open-source framework for large-scale reinforcement learning, with 2,100+ stars on GitHub — fully asynchronous RL training built to teach agents on clusters of a thousand GPUs or more.

Overview

Reinforcement learning for LLM agents has a systems problem: generation and training fight over the same GPUs, synchronous pipelines idle expensively, and most frameworks buckle long before frontier scale. prime-rl, under active development in 2026, is designed around the opposite choices. Training is fully asynchronous for high-throughput agentic RL, built on FSDP2 for training and vLLM for inference, with FP8 inference and parallelism strategies — expert, context, pipeline and data — aimed at trillion-parameter mixture-of-experts models on 1,000+ GPU clusters. It integrates natively with the verifiers library and Prime Intellect’s Environments Hub, including built-in support for SWE (software engineering) and agentic environments, so the environments your agents learn in are first-class parts of the stack rather than afterthoughts. The maintainers describe the codebase as deliberately easy to use and hackable despite the scale it targets.

Key Features

• Fully asynchronous RL – generation and training decoupled for high-throughput agentic learning.
• Frontier scale – designed for 1T+ parameter MoE models on 1,000+ GPU clusters.
• Modern training stack – FSDP2 training, vLLM inference, FP8 and PD disaggregation.
• Environments built in – native verifiers integration and Environments Hub, including SWE environments.
• Hackable by design – readable framework code meant to be modified, not just configured.
• Actively developed – Apache 2.0 licensed with continuous releases from Prime Intellect.

💳 Pricing

prime-rl is free and open source under the Apache License 2.0 for teams running their own clusters. Prime Intellect also operates a hosted compute and environments platform with its own pricing — check the official site if you prefer not to run the hardware.

Who Is It For

AI research labs and ML platform teams training LLM agents with reinforcement learning at serious scale, especially on agentic and software-engineering environments.