# TokenPrint

An interactive 3D visualizer that renders a real transformer's forward pass, showing attention heads, RoPE rotations and the residual stream update live as a small model like Qwen2.5-0.5B or a local GGUF file runs.

TokenPrint is a 3D inspector wired into a real language model forward pass rather than an illustration of one: every visualized value is traced back to the model's own parameters, an actual forward pass, or the GGUF binary header. It can load Qwen2.5-0.5B-Instruct live, or a local .gguf file for models such as Llama 2/3, Gemma, Mistral/Mixtral, DeepSeek/MoE or GPT-2/Pythia, parsed client-side with nothing uploaded.

It offers four modes, from an architecture overview to a generation view, a walkthrough and a debugger with breakpoints and activation ablation, and renders grouped-query attention as real clustered geometry rather than a diagram. Clicking a concept such as RoPE, SwiGLU or RMSNorm links to its original paper. A build-time check fails CI if a placeholder or randomly generated value ever appears in the visualization.

Website: https://tokenprint.in/