Research
Papers are nice. Reproducible evals are nicer.
Safety
Constitutional training, adversarial testing, and capability maps for Azno Mini and Max. We track jailbreak resistance, tool misuse, and long-context failure modes.
Scaling
Azno 1.0 Mini proves you can ship a fast daily model without pretending it is frontier. Azno 1.5 Max pushes reasoning, code, and multi-step agents.
Evals
SWE-bench, terminal tasks, retrieval over long context, and agent loops under production-ish constraints. We report deltas between checkpoints, not just leaderboard screenshots.
Interpretability
We invest in understanding failure modes before we invest in launch-day adjectives.