White Circle publishes the harness behind Halo's 2.3x TRL throughput claim

Halo's trainers are subclasses of Hugging Face or TRL trainers, the parallelism split is set in YAML, and saved weights come back as ordinary SafeTensors that from_pretrained will load ^[7]. White Circle's launch paper sets that against systems such as Megatron, which it says demand a model port, conversion tooling and a separate checkpoint format ^[6]. Onboarding a model family is a wrapper of roughly 100 lines, and the company's published expert-parallelism example runs 127 lines against hundreds or thousands of lines across the Megatron bridge integrations it cites ^[8].

The throughput number comes with conditions. White Circle ran gpt-oss-20b in bf16 on a single eight-GPU B300 node using synthetic, fixed-length sequences, and gave its TRL baseline the same FlashAttention, Liger, fused cross-entropy and grouped-GEMM optimizations ^[11]. At 4,096 tokens and batch size one, Halo reported 9,009 tokens per second per GPU against 3,885 ^[12]. That is 2.32x ^[21], the bottom of the 2.3x to 2.8x band the company advertises ^[3]. Peak memory on a matched fully sharded configuration was 29 GB against 48 GB ^[13], about 40 percent less ^[22].

For that multiple to land on your cluster you would need a mixture-of-experts model of about that size in bf16, that many accelerators of that generation, and sequences already uniform in length ^[11]. White Circle's own numbers show how fast the advantage moves: on Qwen3-30B-A3B it reported anything from a near tie to 1.8x, and at a 256,000-token context length the lead over TRL narrowed to 1.3x ^[14][15].

The test harness is in the repository, so the comparison can be inspected and rerun ^[4].

Halo has no pipeline parallelism and no elastic orchestration, and a resumed job must use the same GPU count it started with ^[16]. On preemptible capacity that means waiting for the same node shape to come back rather than restarting smaller. It also lacks a hosted interface, experiment tracking and hyperparameter search ^[17], so existing tooling stays in place.

The end-to-end demonstration is Z.ai's GLM-4.7-Flash, fine-tuned on 177 million tokens of agentic traces. White Circle says the resulting GLM-4.7-Flash-Coder raised its resolved rate on Nebius's SWE-rebench-V2 from 33.15% to 41.65% ^[18], a gain of 8.5 percentage points ^[23]. The figure covers the dataset and the training run together, not Halo's speed advantage alone ^[18].

The framework also ships fused kernels, a bf16 AdamW optimizer, LoRA support and asynchronous reinforcement learning, with SGLang as the primary rollout engine and vLLM supported ^[9]. White Circle says the same codebase runs LoRA on a 24 GB GPU or scales across nodes of Nvidia B300s ^[10].

White Circle is a Paris-based AI safety startup ^[3]. Denis Shilov founded it in 2025 after finding a prompt that bypassed safeguards across several leading models, which led to private model testing and work in Anthropic's bug-bounty program, according to Fortune's May 12th profile ^[19]. Its product is a runtime control layer that checks model inputs, outputs and agent actions against customer policies ^[19]. The company disclosed an $11 million seed round in May ^[20]. Halo 1.0 has been public since August 20th, 32 days before the announcement thread ^[2][24].