Miles is an enterprise-grade reinforcement learning framework for post-training large and vision language models, co-evolving with the Slime project. The latest v0.1 release highlights production-level capabilities, including day-0 support for Kimi K3 and Inkling. Key technical features include on-policy distillation and native support for Blackwell-era 8-bit and 4-bit precision formats like MXFP8 and NVFP4.
- Miles v0.1 targets enterprise RL post-training, co-evolving with Slime for large-scale model refinement.
- Features day-0 integration with SGLang for Kimi K3 and Inkling multimodal models.
- Supports Blackwell-native 4-bit (NVFP4) and 8-bit (MXFP8) precision for efficient RL.
- Includes on-policy distillation capabilities for improved post-training workflows.