Yaqgu Lab is our ongoing research effort into sequence models that stay light on memory and run on everyday hardware. We are actively prototyping and testing ideas—no claimed benchmarks or production releases yet.
Why we're researching this
Our browser tools already run locally using compact specialized models. Alongside them, we're exploring what a clean, general-purpose base architecture needs to feel native on personal devices.
Standard transformers allocate more memory for every extra token of history. We're investigating recurrent and state-space formulations where working memory stays predictable across long inputs.
Before training anything large, we're studying architecture behavior, optimizer stability, and clean educational datasets at small scale—learning where simpler designs break before scaling up.
Early research involves plenty of dead ends and revisions. Rather than publishing unverified claims, we keep benchmarks private until experiments are reproducible and open to scrutiny.
Current R&D track
Iterating on our experimental Ypru recurrence formulation, evaluating training dynamics, and checking how well compact recurrent states retain long-range details.
Once an experimental checkpoint passes clean sanity checks and ablation tests, we plan to publish the training notes, evaluation scripts, and weights openly.