Map any Hugging Face model — layers, memory, serving cost
A browser tool builds the full graph from a repo ID — no weights pulled — so you can size VRAM before renting a GPU. Anthropic's Console also gets a Playground.
Copy markdown
Paste a repo ID, get the whole architecture
modelmap.cc renders an animated graph of any Hugging Face model — per-layer parameter counts, tensor shapes, and repeated blocks collapsed into one stack with a ×N badge (36 identical layers show once). It builds the model on PyTorch's meta device and runs a fake forward pass, so nothing downloads. Free and in-browser; ~62 points on HN.
Size the GPU before you rent it
Toggle between parameters, compute (MACs), activation memory, and KV cache, then drag a sequence-length slider to watch memory scale. For anyone self-hosting or fine-tuning an open model, that's a serving-cost and VRAM estimate before you pull a multi-terabyte checkpoint. Counts are validated against the safetensors headers via HTTP range requests.
A second viewer for cross-checks
hfviewer.com does a similar per-repo inspection with a different UI — surfaced in the HN thread as an existing alternative. Handy as a sanity check when you're deciding between two models to deploy.
Anthropic's Console gets a real Playground
Workbench is now Playground: it supports every Messages API parameter and ships runnable templates for code execution and web search. It shows the exact SDK request and response, so you can copy working code straight into your app.