# White Circle releases Halo to speed up Hugging Face post-training > Source: > Published: 2026-09-21 20:08:51+00:00 # White Circle releases Halo to speed up Hugging Face post-training **The AI safety startup claims up to 2.8x TRL throughput, while its modified Apache license adds conditions for larger commercial users.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) ยท Published Primary source: [White Circle on X](https://x.com/whitecircle/status/2102087563913609534) ## Why it matters Halo targets the costly gap between single-GPU Hugging Face workflows and heavyweight training systems, while giving White Circle a developer distribution channel beyond its AI safety product. [Denis Shilov (@mixedenn)](https://x.com/mixedenn)'s White Circle released Halo, the internal framework it uses to train and post-train models, giving outside teams a distributed training stack that keeps models and checkpoints in the standard Hugging Face format. [https://x.com/whitecircle/status/2102087563913609534](https://x.com/whitecircle/status/2102087563913609534) White Circle announced the release [in a thread on X](https://x.com/whitecircle/status/2102087563913609534) on September 21st. The repository's changelog dates Halo 1.0's first public release to August 20th, making Monday's announcement the broader launch rather than the code's first appearance. The Paris-based AI safety startup says Halo delivered between 2.3x and 2.8x the throughput of stock Hugging Face TRL in its tests of OpenAI's [gpt-oss-20b](https://runtimewire.com/models/openai/gpt-oss-20b), while consuming less peak GPU memory. Those figures come from White Circle's own benchmarks. The company published its test harness in the [Halo repository](https://github.com/whitecircle/halo), allowing developers to inspect and attempt to reproduce the comparisons. Halo was built by Sergei Bratchikov, Konstantin Korolev, Oleg Smirnov, Mohamed Mekkouri, Dmitrii Kharlapenko and Daniil Oblakov, working with others at White Circle. The group describes the technical gap in its [launch paper](https://whitecircle.com/halo): Hugging Face tools are familiar and comparatively lightweight, while systems such as Megatron demand model ports, conversion tooling and separate checkpoint formats. Halo adds distributed training behavior directly to Hugging Face models. It supports expert, context, tensor and expert-tensor parallelism, with the split configured in YAML. Trainers remain subclasses of Hugging Face or TRL trainers, and saved weights return as ordinary SafeTensors checkpoints that can be loaded with `from_pretrained`. That design targets teams fine-tuning mixture-of-experts models or running reinforcement learning jobs that have outgrown a single GPU, without forcing them to maintain another implementation of each model architecture. White Circle says adding support for a model family can require roughly 100 lines of wrapper code. Its published example for expert parallelism uses 127 lines, compared with hundreds or thousands of lines across cited Megatron bridge integrations. Halo also includes fused kernels, a bf16 AdamW optimizer, LoRA support and asynchronous reinforcement learning. SGLang is the primary rollout engine for online and multi-turn reinforcement learning, while vLLM is also supported. White Circle says the same codebase can run LoRA on a 24 GB GPU or scale across multiple nodes using Nvidia B300 accelerators. The performance claim is narrow enough to inspect. White Circle tested gpt-oss-20b in bf16 on an eight-GPU B300 node using synthetic, fixed-length sequences. Its TRL baseline used the same FlashAttention, Liger, fused cross-entropy and grouped-GEMM optimizations. At a 4,096-token sequence and batch size of one, Halo reported 9,009 tokens per second per GPU against 3,885 for stock TRL, while a matched fully sharded Halo configuration used 29 GB of peak memory against TRL's 48 GB. Results varied by model and workload. On [Qwen3-30B-A3B](https://runtimewire.com/models/qwen/qwen3-30b-a3b), White Circle reported anything from a near tie to a 1.8x advantage. At a 256,000-token context length, Halo's lead over TRL narrowed to 1.3x. The framework does not currently include pipeline parallelism or elastic orchestration, and a resumed job must use the same GPU count with which it started. It also lacks a hosted interface, experiment tracking and hyperparameter search. White Circle used Halo to fine-tune Z.ai's [GLM-4.7-Flash](https://runtimewire.com/models/z-ai/glm-4.7-flash) on 177 million tokens of agentic traces. The resulting GLM-4.7-Flash-Coder model increased its resolved rate on Nebius's SWE-rebench-V2 benchmark from 33.15% to 41.65%, according to White Circle's evaluation. That result measures the combined training run and dataset rather than Halo's speed advantage alone. ### An infrastructure release from a safety startup Shilov founded White Circle in 2025 after finding a prompt that bypassed safeguards across several leading models. The discovery led to private model testing and work in Anthropic's bug-bounty program, according to [Fortune's May 12th profile](https://fortune.com/2026/05/12/exclusive-white-circle-raises-11-million-to-stop-ai-models-from-going-rogue-in-the-workplace/). He built White Circle around a runtime control layer that checks model inputs, outputs and agent actions against policies set by customers. White Circle raised an $11 million seed round disclosed in May. Its backers include OpenAI developer experience chief Romain Huet, Anthropic researcher and OpenAI co-founder Durk Kingma, Mistral co-founder Guillaume Lample and Hugging Face co-founder Thomas Wolf. Halo extends White Circle's reach from controlling deployed AI systems into the infrastructure used to produce specialized models. White Circle says every model it trains now runs through Halo. Publishing that stack can attract researchers and prospective customers before they need White Circle's commercial safety products, while establishing the engineering credentials behind its model-control pitch. The licensing deserves a closer reading than the "open-source" label in White Circle's announcement. Halo uses Apache 2.0 with supplemental terms. Developers can use, modify and self-host Halo, but businesses offering its training functionality as a service need a commercial agreement after generating $20 million in related revenue over 12 months. Organizations with more than $20 million in annual revenue must also add a "Trained with Halo" credit to models produced with the framework. Those conditions preserve White Circle's path to charge large training providers if Halo gains adoption. The immediate test is whether developers adopt another training framework on the strength of company-run benchmarks and Hugging Face compatibility. At launch, the GitHub repository remained early, with 30 stars and no forks, leaving Halo's performance claims open for independent testing.