The Gensyn Public Testnet brings persistent identity to decentralised AI systems and provides a network to track participation, maintain attribution, make payments, coordinate remote execution, verify untrusted operations, log decentralised training runs, crowd-fund large-scale training efforts, and more.

The network is a custom Ethereum Rollup dedicated to machine learning and integrated with off-chain execution, verification, and communication frameworks.

Current applications on the Testnet include:

• RL Swarm: A framework for collaborative reinforcement learning over the internet.
• BlockAssist: An open-source AI assistant that learns from your Minecraft gameplay, training locally and running on your device.
• Judge: A verifiable evaluation framework built on Verde that makes model judgments transparent, reproducible, and cryptographically verifiable.

Phase 0
‍The first phase of the Testnet focuses on tracking participation within RL Swarm, our application for collaborative post-training via reinforcement learning reasoning over the internet. Connecting an RL Swarm node to an on-chain identity provides a consistent view of individual node contributions.

Updates

Update 30/04/2025: The Gensyn Testnet has a new swarm with a more difficult dataset and larger models (up to 72B out-of-the-box). Update your node and run the new swarm using the latest release.

‍Update 25/06/2025: RL Swarm has a new backend: GenRL. This is a complete overhaul to the environment, game, and training logic. RL Swarm can now work with any environment and we’ve incorporated the open source Reasoning Gym library to include >100 out-of-the-box. A prebuilt Docker image is also now available for easier deployment.

Update 06/08/2025: The Gensyn Testnet has a new application, BlockAssist, the open-source AI assistant that learns by watching you play Minecraft. It trains locally and runs on your device, open source and sovereign.

Update 27/08/2025: Introducing Judge, a reproducible evaluation system built on Verde. Judge enables open and cryptographically verifiable evaluation of AI model outputs to eliminate reliance on opaque closed-source judges. It ensures determinism across hardware, traces every operation back to its provenance, and supports scalable verification through refereed delegation. Initial deployments showcase reasoning “prediction market” games, but Judge extends to benchmarks, peer review, and any context where trustworthy evaluation is critical.

‍Get involved

• Participants: Participants: Run a node, train models locally and/or collaboratively, take the future of Artificial Intelligence into your own hands, and track your participation on a global decentralised stage. You can also try BlockAssist, which trains locally as you play Minecraft and doesn’t require a node. It’s a lightweight way to experience decentralised AI, open and entirely under your control.
• Developers: Add passively trained assistance models to your applications, improve them in global swarms across all of your users devices, and remove your reliance on “god models” from external providers.
• ML Researchers: Deploy your own swarms for others to join, train on new datasets, solve new problems, construct new objectives, incentivise participation and explore the space of decentralised AI with an entirely new infrastructure stack.
• Community Members: Stay up to date on progress, provide feedback, and discuss future developments in the Discord.