DI-star is an artificial intelligence platform for StarCraft II from OpenDILab, built around large-scale distributed reinforcement learning training and agents that reached Grandmaster level. The open-source project has around 1,400 stars on GitHub.
StarCraft II is one of the hardest benchmark environments in game AI: partial observability, enormous action spaces, and long-horizon strategy. DI-star provides the full training platform behind a Grandmaster-capable agent — distributed actors generating gameplay, league training in which a population of agents and exploiters co-evolve, and self-play machinery modeled on the approaches proven by DeepMind’s AlphaStar. It builds on OpenDILab’s DI-engine decision intelligence framework and StarCraft II’s official API, and includes the configuration, training scripts, and evaluation tooling needed to reproduce competitive agents rather than toy demos. The platform is valuable both as a research artifact — a working, open implementation of league training at scale — and as an engineering reference for anyone building distributed RL systems. Released under the Apache-2.0 license, its most recent updates were in 2025, and it expects users comfortable with serious GPU compute and RL infrastructure. Few open projects document the full path from environment setup to a championship-caliber game agent, which makes DI-star a reference point for the game AI community well beyond StarCraft II itself.
Free. DI-star is fully open source under the Apache-2.0 license; costs are limited to the compute you train on.
Reinforcement learning researchers and engineers studying large-scale game AI, league training, and distributed training systems.