A model file has a tensor map.
The 3D Model Weights viewer draws each two-dimensional tensor as a slab. Its two sides scale with the square roots of the tensor dimensions. Layers run left to right. Click a slab and the viewer selects the matching subcomponent across the model.
That makes repeated attention and feed-forward blocks readable at a glance. It also exposes architectural differences that disappear inside a parameter total. The live page puts the model list, the 3D structure, and the tensor names on one screen. You can move between the picture and the manifest without losing your place.
The picture answers "what is packed where?" It does not answer "will this run well here?"
How it avoids the payload
The fetch script calls Hugging Face's safetensors metadata parser for each model file. It records tensor names, data types, and shapes. It then folds repeated numeric expert paths into a counted pattern. The project says this uses one header request per file and does not download the weight payloads.
That boundary matters. Hugging Face documents safetensors as a header followed by a byte buffer. Header metadata can reveal structure without loading the tensor data. A shape map can therefore be fast and useful while remaining much smaller than the model it describes.
Where the map stops
The project README is plain about the gaps. Its size estimate is parameter count multiplied by data-type bytes. That is file payload, not total inference memory. Color and layer grouping come from tensor-name guesses. Gated repositories and GGUF-only repositories do not appear.
Runtime memory also has caches, temporary buffers, allocator behavior, and engine overhead. Throughput depends on the serving stack, hardware, sequence lengths, and workload. Quality needs held-out tasks. None of those facts live in a tensor header.
Use it before the download
The viewer is most useful as a first inspection. Read the layer count, hidden width, expert repetition, data types, and rough payload. Then decide whether the model deserves a real intake. The next steps are separate: pin the revision, check the license, fetch the exact files, measure runtime memory, replay a workload, and save the result.