White Circle unveils Halo, an open-source post-training framework

Halo: Halo is an open-source distributed post-training framework developed specifically for large models, supporting pre-training, SFT, preference optimization, and asynchronous multi-turn RL. It integrates directly with Hugging Face Transformers models using expert, context, tensor, and expert-tensor parallelism without requiring architecture reimplementation. The framework runs on a single GPU or across clusters and was publicly released as version 1.0.0 in August 2026. Hugging Face: Hugging Face provides the Transformers library and related tools widely used for open-source AI model development and deployment. Halo builds on this ecosystem by wrapping existing Transformers models and ensuring checkpoints remain compatible with standard from_pretrained loading. The news highlights Halo's design to extend beyond stock Hugging Face TRL capabilities for more advanced distributed training scenarios. White Circle: White Circle is a public benefit corporation focused on developing open-source frameworks for training large language and multimodal models. The company maintains and distributes Halo as its primary product, emphasizing ease of use with existing model ecosystems. White Circle released Halo to address gaps in post-training tools for models that exceed single-framework limits while remaining compatible with standard formats. Licensing: Halo is released under a modified Apache 2.0 license that permits broad self-hosting, modification, and commercial use of trained models with specific conditions for third-party training services. Scalability: The framework supports configuration of multiple parallelism types independently and scales training jobs from single-GPU setups to multi-node clusters using the same codebase and PyTorch-based distributed primitives. Compatibility: Halo preserves native Hugging Face model formats and checkpoint standards so that trained models load seamlessly with existing workflows and from_pretrained calls.