modelmap

Paste a Hugging Face model id and modelmap draws the network as a living map, without downloading any weights. Click a module to inspect it or open its inner steps, and the address bar follows, so a URL reproduces the exact view you are looking at. Two models can be put side by side, where paired modules show both sides and differences come up in amber.

Beyond the diagram it will replay a forward pass, sending a token rippling through the layers, and project each layer's hidden state straight to vocabulary scores so the prediction is visible as it forms. Small models run in the browser: type a prompt, watch real attention and next-token probabilities, zero a head's output and re-run to see what it was contributing. A serving panel does the arithmetic on whether a model fits your GPUs, how to split it, how large the KV cache gets and how many experts an MoE routes per token. Gated repositories need a Hugging Face token, and local checkpoints can be opened directly.