togethercomputer/hybridagent, published on GitHub, is Together's hybrid reasoning agent repository, created on October 9, 2026. Within a single generation run it moves back and forth between latent reasoning (thinking that proceeds only through the model's internal numeric state) and text-based interaction, presenting an agent design that combines a reasoning mode with a non-reasoning mode. It is written in Python and distributed under the MIT license, with 328 stars recorded. No official release has been announced yet.
The README explains that this flow is implemented through a component called HybridLink, which connects the hidden state of the final layer to the input embedding space. Latent states and tool observations are held in task memory, and latent reasoning surfaces as human-readable pseudo text. The repository contains a generation engine and tools, supervised fine-tuning and reinforcement learning stages, batch inference and interpreter functionality, and it points to 8B and 14B models and datasets. With public models and training materials linked, the structure runs from a description of the method through to runnable artifacts.
That said, performance advantages outside the scope of the README have not been evaluated. The scores — freshness 8.0, technique 8.0, quality 5.0, momentum 5.9 — do not provide a basis for comparing this repository with other reasoning agents. There is no basis for a direct comparison at the current stage, so it is best treated as a reference design case.