Introducing Halo, the best framework for post-training of open-source models. Halo delivers up to 2.8x the throughput of stock TRL with less peak memory, while models stay in their native HuggingFace format. Star us on GitHub:

Every model we train at White Circle now uses Halo. The same codebase runs LoRA on a 24 GB GPU, multi-node training on B300s, and async RL. Technical blog:

Other training frameworks often require a separate implementation for each model family. In Halo, adding a new model family takes just 100 lines of connecting a wrapper instead of rewriting the model.

One YAML file contains the model, training method, GPU setup and checkpoint settings. One command starts training.

Halo supports data, tensor, context, expert and expert-tensor parallelism. It also includes optimized attention, grouped-GEMM and fused loss kernels.

Halo is not limited to supervised fine-tuning. It supports async RL and training with external environments. We are excited to partner with the @sgl_project team to make it the primary engine for rollouts.

On @OpenAI gpt-oss-20b, Halo delivered 2.3–2.8x the throughput of stock TRL with less peak memory. Both runs used the same FlashAttention, Liger, fused cross-entropy and grouped-GEMM optimizations.

We partnered with @liquidai to add native Halo support for LFM2.5-8B-A1B. On one B300, Halo was up to 20% faster than next best open-source framework.

We also used Halo to fine-tune @Zai_org GLM-4.7-Flash on 177M tokens of agentic traces. The resulting model improved SWE-rebench-V2 by @nebiusai from 33% to 42%. Halo reached up to 1.63× TRL throughput on the same model precision and data. Model and write-up:

Thanks for reading to the end! Star us on GitHub: Read more:

Open source is the way, and it's only going to get more and more relevant in the coming time! :)

True

insane way to kick off the week -- congrats @whitecircle team

Lets go!

crazy reading through the docs

love that this is open source, congrats team ❤️

Congrats on the awesome work, team! Good to see training infra being open source 👏 Starred the repo to contribute to the growth.

thanks!

great idea you've got my star, guys

took me one read to understand exactly who this is for. that almost never happens with infra launches. congrats team!!! starred 👀

Peak improvement, I will try it from first hand

so we get Grok 4.7 along with this today? nice.

Crazy day!

Congrats guys!

This is huge !!

haloshi framework

wait what this is incredible

ty Brian! show the repo some love ⭐

so excited!!!

🥰

Congrats team!

Great work! You got my star ⭐️

love that this is open source, congrats!

I have always wanted to train a custom math model, let’s see how this goes!

Keep us posted, any feedback is welcome

love seeing more training infrastructure get open-sourced. congrats on the release!

Congrats! Impressive what you delivered there 👀

@sarapenrique This is fabulous

Uh training my own models for my products would be pretty neat. I need better hardware 👀

💚

oh this is sick actually, well cooked

banger stuff 💥

insane start of the week congrats team!

wait… 2.8x faster than stock trl ??

congrats to the team!

congrats!

Thank you

congrats! def keeping an eye on this one.

The interesting part is keeping models in native HuggingFace format while getting that throughput boost.

Worked hard on this

LFG 🫡🙌

Really cool. Really unique and impressive numbers :•)

The only limit now is the hardware. Everyone can now train or finetune a model without rewriting it for the training phase. Awesome!

thanks Marco!

Great work !

Congrats on the launch team! Starred.

happy launch day guys"