Jittor is a new deep learning framework proposed by the visual media intelligent computing team of the Beijing National Research Center for Information Science and Technology. It uses meta operators to express the computing units of neural networks, organizes them into a unified computational graph for optimization, and compiles them just in time. Jittor was officially released under an open source license on March 20. After it was released on GitHub, a well-known open-source platform, it was met with immediate good comments from deep learning developers. Major news outlets and technology we-media outlets also reported on Jittor's release.

Jittor designs a new dynamic compilation framework from the bottom, which ensures the separation of implementation and optimization, and greatly improves the flexibility, scalability and portability of application development. Jittor has many advanced features, such as support for automatic differentiation, dynamic graphs, synchronous and asynchronous interfaces, unified memory management, and cross-iteration fusion. Compared to other frameworks in the same category, Jittor delivers a 10-50% improvement in multi-task reasoning and training speed with the same convergence precision.