Chainer is a Python deep learning framework from Preferred Networks, historically important as the framework that popularized the define-by-run approach — building the computation graph dynamically as the code executes, the design later mainstreamed by PyTorch. It has about 5,900 stars on GitHub and is released under the MIT license.
Chainer offers a NumPy-like API for defining models, with GPU acceleration through CuPy and a fast C++ core called ChainerX for supported operations. At its peak it powered research and production systems across Japanese industry and academia, with a rich set of model examples. The project has stated openly since its v7 release that it is in a maintenance phase: development is limited to bug fixes and maintenance rather than new features, and the last repository activity was in 2023. It remains a solid, well-documented framework for the many codebases built on it.
Chainer’s place in deep learning history is secure: released in 2015, it was the first major framework built around define-by-run, and its design directly influenced the dynamic-graph direction the whole field eventually took. Preferred Networks used it for demanding industrial work, including inspection systems and robotics, which kept the framework battle-tested even as the community’s center of gravity moved elsewhere. The associated CuPy library — born as Chainer’s GPU array backend — very much lives on and remains widely used, even though Chainer itself now receives only maintenance attention.
Free. Chainer is open source under the MIT license.
Teams maintaining Chainer-based systems, researchers reproducing results from Chainer-era papers, and anyone studying the define-by-run design that shaped modern frameworks.