White Circle open-sources Halo, a faster Hugging Face trainer

Paris startup White Circle open-sourced Halo, a distributed post-training framework that keeps models in native Hugging Face format and claims up to 2.8x TRL throughput.

In this article

What it changes about post-training
The benchmark, and what it does not tell you
A licensing detail worth reading before adopting

White Circle, a Paris-based AI-safety startup, open-sourced Halo on 21 September 2026 — a distributed training framework built to speed up post-training of Hugging Face models while keeping them in the standard format the rest of the ecosystem expects. The company built Halo internally before releasing it; the repository's own history shows an initial release a month earlier, on 20 August 2026.

A silver tuning fork lying across an open page of printed sheet music, with soft focus around it — Photo: Lukasz Kobus, European Commission (CC BY 4.0), Wikimedia Commons.

What it changes about post-training

Halo wraps existing Hugging Face model definitions rather than replacing them: checkpoints still load through the familiar from_pretrained call, and stay as standard SafeTensors files. White Circle says supporting a new model family typically takes well under 150 lines of integration code. Under the hood, the framework adds expert, context, tensor and combined expert-tensor parallelism, along with LoRA/QLoRA fine-tuning and asynchronous multi-turn reinforcement learning.

The benchmark, and what it does not tell you

White Circle reports Halo delivers 2.3 to 2.8 times the throughput of stock Hugging Face TRL in its own tests on OpenAI's gpt-oss-20b model, using the same attention and kernel optimizations in both systems so the comparison isolates the trainer and optimizer. These are the company's own benchmarks, not an independently reproduced result, and outside commentary has noted the advantage narrows at very long context lengths. Teams evaluating Halo should treat the headline multiplier as a starting point tied to one model and one hardware setup, not a guarantee for a different workload.

A licensing detail worth reading before adopting

Halo ships under a modified Apache license that adds commercial conditions for larger organizations or higher-revenue users, rather than plain Apache 2.0. Teams planning production use, especially at scale, should read the license text directly rather than assume unrestricted open-source terms.

For MSPs and product teams already running Hugging Face-based training pipelines, Halo is worth a pilot on a representative workload before any throughput claim is taken at face value.

Sources: GitHub — whitecircle/halo, RuntimeWire — White Circle releases Halo.