RLX
A reinforcement learning framework built on Apple's MLX array framework, with single-file implementations of algorithms such as DQN, PPO and SAC that run end-to-end on device, including the environments, for over a million steps per second on Apple silicon.
RLX runs reinforcement learning environments, replay buffers and the learner entirely on MLX, so nothing round-trips between the device and host during training. Every update step is wrapped in MLX's compile function to fuse the training graph, and the correct backend, Metal on Apple silicon or CUDA on Linux, is selected automatically.
Each algorithm, including DQN, REINFORCE, A2C, PPO, SAC and TD3, is a single, self-contained file in the style of CleanRL. Bundled MLX-native environments cover classic control tasks such as CartPole, MountainCar, Acrobot and Pendulum, plus branchless ports of the MinAtar suite validated against the reference package, and an EnvPool adapter lets other pre-vectorized environments plug into the same interface.