Google releases TensorFlow as open-source software

Google released TensorFlow today under the Apache 2.0 open-source license. The software expresses machine-learning work as data-flow graphs, allowing the same model description to operate across CPUs, graphics processors, desktops, servers and mobile devices.

The system grew from Google’s internal DistBelief infrastructure and is already used in products and research. The initial release includes code, documentation and tutorials, though Google notes that it does not yet provide the company’s large-scale distributed training environment.

Opening the framework could give researchers a common language for reproducing experiments and moving models into real applications. It also places Google’s preferred tools at the center of a fast-growing technical field. TensorFlow’s real test will be whether outside contributors can understand, extend and trust software originally shaped around one company’s needs and infrastructure.