Halo is an open-source framework for pre-training and post-training large language and multimodal models, released by the Paris AI safety startup White Circle. It targets teams who have outgrown single-GPU Hugging Face workflows but want to avoid the conversion tooling and custom checkpoint formats that come with heavyweight training systems. Halo trains Hugging Face models directly. Trainers subclass Hugging Face or TRL trainers, checkpoints remain standard SafeTensors, and from_pretrained continues to work.
The launch matters because distributed post-training has largely meant either patching TRL and Accelerate past their limits or moving to Megatron-style stacks that require rewriting the model. Halo takes a different path by adding parallelism to existing Hugging Face models in place, with EP wrapping MoE blocks, CP wrapping attention, and TP/ETP sharding weights, built on PyTorch FSDP2, DTensor, and DeviceMesh, with DeepEP handling all-to-all communication for Expert Parallelism . On the performance side, on 8x B300, Halo delivers up to roughly 2.8x the training throughput of stock TRL (2.7x at 25% less peak memory when both sides shard ZeRO-3), with larger margins over the other frameworks benchmarked . Runs are configured through readable YAML and started with a single command inside the provided Docker image.
Halo is the training stack that White Circle, founded by Denis Shilov after his viral 2024 universal jailbreak exposed safety flaws in major AI models , has been using internally alongside its runtime control platform for enterprise AI. The company has processed more than 1 billion API requests and works with customers including Lovable and two of the world's largest digital banks , and recently raised an $11 million seed round with backing from figures including Romain Huet of OpenAI, Durk Kingma of Anthropic, Guillaume Lample of Mistral, Thomas Wolf of Hugging Face, Olivier Pomel of Datadog, François Chollet of Keras, and David Cramer of Sentry . Releasing Halo extends that work upstream into the training loop where model behavior is actually shaped.
Comments (11)
Sign in to join the discussion.
N
Nadia Burek22d ago
2.8x throughput is a spicy number but the tweet buries it under 'post-training framework'. Lead with the benchmark next time, the video literally shows the graph.
K
Kwabena O.22d ago
Finally a framework that doesn't demand I convert my weights into some cursed proprietary shard format. Staying in native HF is the actual headline here.
P
Priya V.22d ago
Two people, three, or is there a whole team behind Halo? Curious what the shipping crew size looks like on something this infra-heavy.
M
miren22d ago
Wait, White Circle does model post-training now? I thought you were the guardrails people. Pivot or expansion?
D
Deepa Mathur22d ago
Would love to hear the story behind the pivot from behavior control to training infra for a piece I'm working on. DMs open whenever you have 15 minutes.
R
Ravi Prakash22d ago
The tagline on the site says 'control user and model behavior in real time' and the launch is a training framework. Pick a lane or explain the throughline in the thread.
S
Sonia F.22d ago
Star count is going to spike from this tweet and then plateau on Sunday, watch. The real signal is whether anyone opens a non-trivial issue by Wednesday.
T
Tomasz K.22d ago
Halo, huh. Bold choice naming your framework after the thing everyone's been shipping since 2018. Hope the SEO team stretched first.
A
Amara Ogunleye22d ago
Any chance of a webhook or callback hook for eval runs mid-training? Would love to pipe intermediate checkpoints into our own harness without polling.
L
Leo Zhu22d ago
Peak memory reduction is the part I actually care about. Throughput I can buy with more GPUs, VRAM ceilings I cannot.
F
Finn Holgersen22d ago
The launch video pacing is solid but that voiceover mix is way hotter than the background music. Almost clipped in the second cut.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
White Circle
Control user and model behavior in real time
Control user and model behavior in real time
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Halo is an open-source framework for pre-training and post-training large language and multimodal models, released by the Paris AI safety startup White Circle. It targets teams who have outgrown single-GPU Hugging Face workflows but want to avoid the conversion tooling and custom checkpoint formats that come with heavyweight training systems. Halo trains Hugging Face models directly. Trainers subclass Hugging Face or TRL trainers, checkpoints remain standard SafeTensors, and from_pretrained continues to work. The launch matters because distributed post-training has largely meant either patching TRL and Accelerate past their limits or moving to Megatron-style stacks that require rewriting the model. Halo takes a different path by adding parallelism to existing Hugging Face models in place, with EP wrapping MoE blocks, CP wrapping attention, and TP/ETP sharding weights, built on PyTorch FSDP2, DTensor, and DeviceMesh, with DeepEP handling all-to-all communication for Expert Parallelism . On the performance side, on 8x B300, Halo delivers up to roughly 2.8x the training throughput of stock TRL (2.7x at 25% less peak memory when both sides shard ZeRO-3), with larger margins over the other frameworks benchmarked . Runs are configured through readable YAML and started with a single command inside the provided Docker image. Halo is the training stack that White Circle, founded by Denis Shilov after his viral 2024 universal jailbreak exposed safety flaws in major AI models , has been using internally alongside its runtime control platform for enterprise AI. The company has processed more than 1 billion API requests and works with customers including Lovable and two of the world's largest digital banks , and recently raised an $11 million seed round with backing from figures including Romain Huet of OpenAI, Durk Kingma of Anthropic, Guillaume Lample of Mistral, Thomas Wolf of Hugging Face, Olivier Pomel of Datadog, François Chollet of Keras, and David Cramer of Sentry . Releasing Halo extends that work upstream into the training loop where model behavior is actually shaped.
Comments (11)
2.8x throughput is a spicy number but the tweet buries it under 'post-training framework'. Lead with the benchmark next time, the video literally shows the graph.
Finally a framework that doesn't demand I convert my weights into some cursed proprietary shard format. Staying in native HF is the actual headline here.
Two people, three, or is there a whole team behind Halo? Curious what the shipping crew size looks like on something this infra-heavy.
Wait, White Circle does model post-training now? I thought you were the guardrails people. Pivot or expansion?
Would love to hear the story behind the pivot from behavior control to training infra for a piece I'm working on. DMs open whenever you have 15 minutes.
The tagline on the site says 'control user and model behavior in real time' and the launch is a training framework. Pick a lane or explain the throughline in the thread.
Star count is going to spike from this tweet and then plateau on Sunday, watch. The real signal is whether anyone opens a non-trivial issue by Wednesday.
Halo, huh. Bold choice naming your framework after the thing everyone's been shipping since 2018. Hope the SEO team stretched first.
Any chance of a webhook or callback hook for eval runs mid-training? Would love to pipe intermediate checkpoints into our own harness without polling.
Peak memory reduction is the part I actually care about. Throughput I can buy with more GPUs, VRAM ceilings I cannot.
The launch video pacing is solid but that voiceover mix is way hotter than the background music. Almost clipped in the second cut.