Runtime
Reusable C++ and CUDA runtime for weight-shared looped transformers.
TinyLoop Framework
A runtime for low-bit, weight-shared transformer inference.
TinyLoop is a focused framework for looped transformer systems: packaged, documented, and structured for teams working on weight-sharing architectures rather than generic model serving.
TinyLoop keeps the runtime opinionated: loop reuse, low-bit weights, explicit model format boundaries, and a deployment surface built around that architecture instead of a general transformer stack.
Highlights
Interface
CMake package, CLI, and Python binding with last-token scoring support.
Artifacts
Versioned .tinyloop format with explicit loader validation and deployment docs.
Read next