etude labs Join the pilot

Blog

Ideas and evidence behind continually improving inference.

Founder theses, benchmark methodology, and results from building the model-improvement layer behind the endpoint.

Introducing Ballade-1

We post-trained Qwen 3.8 Max on 1,000 verifiable weight-editing tasks and benchmarked its ability to improve Gemma 3 4B through direct weight updates.

The Inference-Time Learning Gap

What Ballade-1 shows—and when a learned weight update can pay for itself.