White Circle released Halo 1.0 on September 21, 2026, opening the training and post-training framework it uses for its own models.
In the Halo launch post, the company describes Halo as a distributed layer for Hugging Face model workflows that keeps standard SafeTensors checkpoints. Supported methods include supervised fine-tuning, preference training, reward modeling, GRPO, and distillation. Halo also supports asynchronous reinforcement learning with SGLang or vLLM rollouts, plus sandboxed tool use during training.
That gives teams already built around Hugging Face a way to expand into multi-node post-training without moving to a heavier stack or changing checkpoint formats.
There are still caveats. RuntimeWire says the published benchmark results come from White Circle's own testing and may vary by workload. It also reports supplemental license terms that add conditions for larger commercial users, so legal review may matter alongside benchmark validation.