what is token-print fr?

TokenPrint is a browser tool that lets you see the inner workings of a language model in interactive 3D. You can load a live small model or drop in your own model file, and every number shown is real, either parsed directly from the file or produced by an actual run of the model, nothing is faked or illustrative. The tool has four modes. Architecture shows a 3D point cloud of every real tensor in the model along with an overview of its size and structure, plus a searchable list and a view for comparing different quantized versions of the same model. Generation streams a real token by token run of the model over a live connection, drawing a distinct 3D shape for each part of the model such as one blade per attention head, while a side panel shows the matching math formula and live probability scores. Walkthrough is a guided, chaptered tour covering topics from tokenization through the final output layer, timed to real data rather than placeholder animations, and includes prediction games using sample prompts in different languages. Debugger is a more technical dashboard with breakpoints, a flame graph of operations, detection of unusual activity, and the ability to compare a model before and after specific parts are disabled. The visuals are built directly from real numbers: the count of shapes for attention matches the real number of attention heads, and the size of certain shapes matches real internal ratios of the model. The project runs a Python and FastAPI backend for the live model and generation, alongside a Next.js frontend using React Three Fiber for the 3D graphics, and it automatically recovers if the browser's 3D context is lost.