White Circle open-sources Halo, a framework for training open models

White Circle, which builds tools to control user and model behaviour in AI applications, has released its model training framework Halo as open source on GitHub. The company says every model it trains now uses Halo.

What Halo is

Halo is a software framework for training language models and multimodal models that also understand images. Its main focus is post-training: taking an existing open model and fine-tuning it on your own data, aligning it with preferences, or improving it at a specific task with reinforcement learning. It also supports pre-training from scratch.

Halo's key feature is that it works directly with Hugging Face models. The trained model stays in Hugging Face's own format and loads the usual way after training. Supporting a new model family takes a small piece of integration code, with no need to rewrite the model.

What it can do

• Same code at every scale: LoRA fine-tuning of a small model on one GPU and training large mixture-of-experts (MoE) models across multi-node clusters use the same codebase.
• Training methods: Pre-training, supervised fine-tuning (SFT), preference optimisation (such as DPO), distillation and reinforcement learning share the same infrastructure.
• Agent training: In multi-turn reinforcement learning, where the model works through tasks in an environment, rollouts run on a vLLM or SGLang server alongside training.
• Ready for coding agents: The repository ships skills that Claude Code and Codex can pick up, for launching jobs, debugging and adding models. The docs come in two sets: short guides for humans and a detailed technical reference written for coding agents.

How fast it is

According to White Circle's own benchmarks on a server with eight NVIDIA B300 GPUs, Halo processes up to 2.8x more data than a standard setup of TRL, Hugging Face's widely used training library, while using less memory. The company says training results stay in line with TRL. The methodology is in the repository's performance document.

How to use it

Halo comes with ready Docker images for NVIDIA's Blackwell and Hopper GPUs, which can be pulled without an account. A training run starts with a single command using one of the example config files. The repository includes ready examples for different models and methods.

Things to keep in mind

• The license adds supplemental terms on top of Apache 2.0. Training your own models and using them internally is unrestricted. But companies that sell Halo to others as a training service and earn more than $20 million a year from it must sign a separate agreement with White Circle. Companies with annual revenue above $20 million that publish a model trained with Halo must add "Trained with Halo" to the model card.
• The speed comparisons were run on data-centre GPUs. For a fuller picture, follow evaluations from independent researchers as well.
• Training large models needs multi-GPU data-centre hardware. The single-GPU option is for LoRA fine-tuning of small models.

Sources: GitHub: whitecircle/halo, White Circle blog: Halo, Frontier-Lab Training for Everyone, Halo license.