areal-project/AReaL

Health v2 · maintenance only

Adoption not part of the score

5696 stars · 583 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

AReaL is a large-scale asynchronous reinforcement learning system that bridges foundation model training with agent-based applications, supporting algorithms like PPO, GRPO, and DAPO on backends such as Megatron, FSDP, and SGLang. Version 2.0 refactors it into a microservice architecture with independent training, inference, agent, and weight-update services for scalable agentic RL.

Use cases

• train reasoning LLMs with reinforcement learning
• run agentic RL post-training for LLM agents
• fine-tune models with GRPO or PPO
• train agents with online RL against black-box APIs
• scale asynchronous RL training across GPU clusters
• post-train multi-turn tool-using agents with verifiable rewards

When to choose

• you need scalable asynchronous RL training for large language models
• you want to train reasoning, coding, or search agents with RL
• you need to mix RL algorithms with popular training and inference backends
• you want online RL training against black-box agent applications

When to avoid

• you only need simple supervised fine-tuning without RL
• you lack multi-GPU or cluster infrastructure
• you need a lightweight single-GPU hobby setup

Facets

framework · maturity active

llm-training reinforcement-learning agent-framework machine-learning gpu-computing llm-inference reinforcement-learning large-language-models machine-learning deep-learning gpu-computing python cloud agentic-rl asynchronous-training grpo ppo rlhf post-training sglang megatron fsdp mlsys ai-agents gpu linux docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="areal-project/AReaL")