etude labs
Continual learning reimagined.
Inference endpoints for open-weight models
that learn while your agents work.
The work
Useful experience should change the model itself, not a markdown file.
We’re building inference endpoints for long-running agents. Customers call one API; we handle serving, evaluation, and the promotion of verified model updates. Targeted weight edits and post-training turn useful feedback into changes behind that endpoint.
Inference that keeps learning
We post-train our own model to edit the weights of open models during inference.
As your agents run, it reads their past trajectories and the problems in front of them right now, and patches the model so the next attempt goes better. Every edit is checked against held-out evaluations before it serves traffic.
Read the Ballade-1 reportBlog
What we are building, why models should learn mid-task, and how we train a model to edit another model's weights.