Orbit
A reinforcement-learning post-training framework that keeps trillion-parameter language models at low-precision deployment format and trains only a small BF16 adapter, fitting frontier-scale RL onto a single node.
Orbit holds a model's base weights at deployment precision, such as INT4, FP4 or FP8, and puts gradients only on a small BF16 OFT or LoRA adapter, so training and serving share the same kernels and quantization scheme with no precision gap between them.
The maintainers report using it to run stable, end-to-end reinforcement learning on trillion-class models, including Kimi-K2.6 and DeepSeek V4-Pro, entirely on a single eight-GPU node, with parity preflight gates checked between the Megatron and SGLang backends. PEFT KL launchers compute reference log-probabilities in-model rather than through separate reference workers, with asynchronous double-buffering of the adapter.
It requires Python 3.12, CUDA 13.2 and PyTorch 2.11, alongside sibling checkouts of the Megatron-Bridge and SGLang backends, and is licensed under Apache-2.0.