ART

What it is

ART is a reinforcement learning framework that enables the training of multi-step agents for real-world tasks. It supports Qwen2.5, Qwen3, Llama, and more, providing on-the-job training capabilities.

ART stands out by providing a specialized framework for multi-step reinforcement learning, focusing on real-world task training and leveraging GRPO.

When it fits — and when it doesn’t

Known weaknesses

Each one links to the public issue, changelog or docs page it came from, where we have one.

1. 01

   Performance issues at scale

   Known to slow down significantly when training large multi-step agents on complex tasks

What it does well

• Supports training multi-step agents for real-world tasks

• Uses GRPO (Generalized Reward Prediction Objective) for reinforcement learning

• Compatible with Qwen2.5, Qwen3, Llama, and more

Pricing

Free tier

YesOpen source — free to use

Paid from

$0

Model

Flat rate

Enterprise plan

No

Pricing last confirmed Sep 21, 2026.