RL Swarm

RL Swarm is Gensyn’s fully open-source framework for running reinforcement-learning training swarms over the internet. Volunteers and teams contribute compute from ordinary machines, and the swarm collaboratively trains models on RL tasks — most famously multi-agent “animal” games and reasoning environments — coordinating progress without a central training cluster. The project has around 1,700 stars and its code was last updated in 2026.

Overview

Gensyn builds decentralised machine-learning infrastructure, and RL Swarm is its public testbed for collaborative training. Each node runs RL rollouts locally, shares what it learns with the swarm, and periodically syncs with peers, so the collective model improves faster than any single participant could manage. The repo includes the node software, the training environments and the coordination logic needed to join or run a swarm.

Key Features

• Decentralised RL training across many independently operated machines
• Ready-made swarm environments, including multi-agent and reasoning tasks
• Peer coordination and periodic model synchronisation built in
• Runs on consumer hardware — no datacenter required to participate
• Open implementation of Gensyn’s collaborative-training research
• MIT licensed and actively developed

💳 Pricing

Free. RL Swarm is open source under the MIT license; you run the node software on your own machines at no cost.

Who Is It For

RL researchers interested in distributed and collaborative training, decentralised-AI enthusiasts, and hobbyists who want to contribute spare compute to shared model training.

Gensyn positions the swarm as an experiment in open, collaborative model training: participation is open to anyone with a machine to contribute, and the shared environments give the network a common benchmark for measuring collective progress over time.