uTensor

Why we included this project

uTensor targets a world most ML tooling ignores: getting inference to run on Cortex-M microcontrollers, where a model has to fit into tight flash and RAM budgets. The workflow is straightforward if you already train in TensorFlow. It translates a trained model into generated C++11 code you drop straight into an embedded project, so the device runs inference on its own with nothing in the cloud. The core runtime stays around 2 KB, and the reworked API leans on compile-time checks, explicit memory ownership through allocators, and a graph you can step through in GDB rather than treating as a black box. If you want to see how a real inference runtime.