Archai is Microsoft’s open-source platform for Neural Architecture Search (NAS), hosted on GitHub with about 490 stars under the MIT license and updates in 2026. Its stated goal is to make NAS fast, reproducible, and modular, so researchers can generate efficient deep networks without rebuilding the search machinery each time.
Neural architecture search automates the trial-and-error of designing neural networks, but experiments are notoriously hard to reproduce and compare. Archai provides a common framework in which search algorithms, search spaces, and evaluation pipelines can be mixed and rerun consistently. It implements well-known NAS approaches such as DARTS and PetriDish, and also covers adjacent ground including hyperparameter optimization and model compression. The platform is built on PyTorch, installs from PyPI with a single pip command, and is documented on its project site with task guides for common scenarios.
Architecture search is compute-hungry, so being able to reproduce a published result before spending GPU days on it matters. Archai’s shared abstractions for search spaces, algorithms, and trainers are designed to make that verification routine, and to let a better search algorithm be dropped into an existing experiment without rewriting it. The documentation groups guidance by task, so practitioners can start from a concrete goal, such as compressing a model, rather than from the theory.
Archai is free and open source under the MIT license. There are no paid tiers; you run searches on your own hardware or cloud compute.
It is aimed at ML researchers and engineers working on efficient model design, especially those who need to compare NAS algorithms fairly or search for compact architectures for constrained deployment targets.