OpenPipe ART
An open-source reinforcement learning framework for multi-step AI agents, with local GPU training, cloud options, and support for Qwen, GPT-OSS and Llama.
OpenPipe ART is a Python framework for developers who want to train multi-step AI agents on the tasks their applications actually perform. It uses reinforcement learning through GRPO (Group Relative Policy Optimization) to improve model behavior based on experience and task rewards. It's open source under Apache 2.0.
Training can run on your own GPU or in a cloud GPU environment. The client works on any machine that runs Python, so your development laptop doesn't need a GPU when training happens elsewhere. ART separates the application client from the training server, letting developers add reinforcement learning to an existing application while choosing where the model runs. W&B Training provides a hosted alternative that manages training and inference infrastructure with autoscaling GPUs.
ART supports model families including Qwen, GPT-OSS and Llama. It uses vLLM for inference and LoRA adapters for training, with broader compatibility across causal language models supported by vLLM, Hugging Face Transformers or Unsloth. Its OpenAI-compatible client connects application workflows to the model as training progresses.
Developers can use ART to address recurring agent mistakes or train for tasks such as research, information retrieval and scripted voice interactions. Training and inference settings are customizable, with defaults aimed at efficiency and stability. Hosted integrations with W&B, Langfuse and OpenPipe support observability and debugging.
