White Circle open-sources Halo post-training framework, claims up to 2.8x TRL throughput

On Sept 22, AI safety company White Circle open-sourced Halo on GitHub, a post-training framework for large language and multimodal models under a modified Apache 2.0 license. The company claims up to 2.8x throughput over native TRL with lower peak memory, and that model weights stay in native HuggingFace format. The release was boosted by a repost from Hugging Face CEO Clem Delangue; the repo had roughly 282 stars early on.

Confirmed

Open-sourced on GitHub under a modified Apache 2.0 license (m4, m6, m8)
Up to 2.8x throughput vs native TRL with lower peak memory, per official claims (m1, m2)
Positioned between HuggingFace TRL and Megatron: for models beyond TRL's comfort zone that don't need the full Megatron stack (m3)
Weights stay in native HuggingFace format, no conversion needed (m2, m7)
Full engineering setup: Docker and docker-compose, with both sglang and vllm deployment options, docker-compose including EF-related config (m4, m6, m8)
Ships with two MoE fine-tuning recipes (m1)
282 stars at open-source launch (m8)
Hugging Face CEO Clem Delangue endorsed it, saying training models is getting easier (especially with TRL and agents) and that those using off-the-shelf models for every task are missing out (m5)

Unconfirmed

The 2.8x throughput and lower memory figures are vendor claims without independent third-party verification

Why it matters

Halo targets the gap between TRL and the full Megatron stack for open-model post-training, with TRL-beating throughput claims, native HF format compatibility, and complete deployment tooling, lowering the barrier to fine-tuning one's own models
The public endorsement from Hugging Face leadership adds community credibility amid the broader industry debate favoring fine-tuned over off-the-shelf models

2026-09-22 ~ 2026-09-22 · 8 related posts

Primary sources

White Circle open-sources Halo, a framework for training LLMs and multimodal models — rohanpaul_ai · 2026-09-22 01:55 (UTC+08:00)
Halo: open-source post-training framework claims 2.8x TRL throughput with lower memory — _akhaliq · 2026-09-22 01:37 (UTC+08:00)
White Circle open-sources Halo, a framework for training LLM and multimodal models — aigclink · 2026-09-22 07:31 (UTC+08:00)

[source] Halo: open-source post-training framework claims 2.8x TRL throughput with lower memory — _akhaliq · 2026-09-22
White Circle open-sources Halo, a training framework for LLMs and multimodal models — prasenx · 2026-09-22
Halo fine-tuning framework launches with 2.8x TRL throughput and MoE recipes — JosephJacks_ · 2026-09-22
Halo post-training framework claims up to 2.8x TRL throughput with lower peak memory — huggingface · 2026-09-22
White Circle open-sources Halo, a distributed post-training framework hitting 2.8x TRL throughput — rohanpaul_ai · 2026-09-22

3 near-duplicate retellings: rohanpaul_ai · aigclink · prasenx