Miles is an enterprise-grade reinforcement learning framework for large-scale model post-training, developed by RadixArk and available on GitHub with about 3,181 stars under the Apache-2.0 license, actively developed in 2026. It began as a fork of the slime framework and continues to co-evolve with it, hardened for production post-training of both language and vision-language models.
Post-training is where base models become useful products, and Miles focuses on doing it at production level: scalable reinforcement learning pipelines for aligning and improving large models after pretraining. The framework sits in the SGLang ecosystem, which shows in its release cadence: Miles and SGLang regularly add day-zero support for major new model releases, so teams can start post-training the newest architectures immediately. A published technical report describes the Miles v0.1 system design, and the project explores post-training for understanding and generating the multimodal world, with a companion project extending the approach to diffusion models. Documentation, quick starts, and a supported-models list are maintained on the official site, and the community coordinates through the SGLang Slack. For organizations, the appeal is a framework exercised in real production runs rather than only in paper experiments, with documentation and community support channels maintained alongside the code.
Miles is free and open source under the Apache-2.0 license. As with any post-training framework, your real cost is GPU infrastructure.
Miles is for AI platform teams and research labs running serious post-training programs, especially those standardizing on SGLang for inference who want training and serving to move in lockstep.