GitHub radar

Netherite: 7,200 Minecraft Worlds on One GPU for RL

An independent researcher's from-scratch C/CUDA Minecraft engine for training AI agents at scale — 7,200 parallel game worlds running simultaneously on a single GPU.

Created by independent developer Infatoshi, Netherite is a from-scratch C/CUDA reimplementation of Minecraft 1.11.2 that is bit-verified against the original Java game — producing identical results frame by frame. Its CUDA-batched reinforcement learning environment runs 7,200 independent game worlds simultaneously on a single GPU, all stepping in lockstep, making large-scale RL agent training dramatically faster.

Why a vibe-coder should care

Reinforcement learning — the training method behind many modern AI capabilities — normally requires massive compute. Netherite compresses that onto a single gaming GPU by running thousands of Minecraft environments in parallel, making large-scale agent research possible without a data center.

Open on GitHub