Halo

Halo is an open source distributed training framework from White Circle, the AI safety company that uses it to train all of its own models. It adds parallel training directly to Hugging Face models, so there is no porting step and no new checkpoint format. Adding a new model family is described as roughly 100 lines of code instead of a full reimplementation. It targets teams that have outgrown stock Hugging Face tooling such as TRL but are not large enough to justify Megatron. According to White Circle's own benchmarks, Halo delivers 2.3 to 2.8 times the throughput of stock TRL on the same hardware at lower peak memory. It brings the parallelism strategies needed for Mixture of Experts models, along with fused kernels, a bf16 optimizer, sequence packing and an async reinforcement learning loop. Beyond plain fine tuning, the write up lists agentic and async RL, multi turn RL, on policy distillation and embeddings learning as supported uses. The audience is machine learning engineers and research teams training or fine tuning large language and multimodal models who want to stay in the Hugging Face ecosystem. The framework was announced in September 2026 and the code is on GitHub. The page offers Get Started and Contact Sales options, but no prices are published. Because the framework is open sourced, it can be used without a license fee, though the vendor does not describe any paid tier.