markab21/rpn-gpt is a project that compiles an exact Forth machine into the weights of Nemotron 3.5 Lightning and post-trains the model to use that machine. The repository description proposes a setup in which the model runs a computational stack machine alongside generating programs in ordinary text. Its focus is not the changes of a particular version, but an experimental design of embedding a computational device inside the model.
Reverse Polish notation (RPN) is a notation that pushes values onto a stack first and places operators after them. In this project, the model executes the programs it writes as part of the generation process. The README explains, with diagrams and per-token examples, how the machine steps are performed using hidden tokens and how the results are written onward, and it also provides execution logs and calculation examples. The setup includes Nemotron 3.5 Lightning, LoRA, a Forth machine computed in weights, training data, and the model in several deployment formats. The repository is written in Python and is MIT licensed.
The core of this approach is connecting text generation and computation inside the model. However, the provided information contains no performance scores or comparison results, and no latest release is listed either. The examples and execution logs therefore show how the design behaves, but they are not sufficient grounds for judging actual computational performance.