Anatomy
Each box is one part of the model, in the order data flows through it. The strip shows the type of every layer. Click a box to see what is inside and where each value comes from.
Press ▶ Play to walk the parts: what goes in, what comes out, and why. Click any part to open it. Switch the lens to see where the weights, compute, ops or precision live.
Every layer, drawn to scale. Click one to open it.
Block width = the real value of the chosen lens. Encoder row on top (if any), decoder row below (if any).
The vocabulary, as moving pictures
The numbers in each card come from the selected model. The last one is yours to play with.
10244096Attention
Feed-forward
KV cache
Tied embedding
Seams
Round 64 weights to fewer levels
Blue dots = original · pink line = after rounding · add the outlier and watch small weights collapse to 0 · switch to one scale per 32 and watch the error drop.
Where it all goes
The selected model in numbers. Copy as Markdown for slides or reports.
What this tool does
Paste a Hugging Face repo id, or drop a .safetensors, .tflite or config.json file. The page reads only the safetensors header: a few tens of kilobytes that list every tensor's name, shape and type. The weights are never downloaded. From that list it finds the repeated layers, labels each part as attention, feed-forward, mixture of experts, SSM or linear attention, convolution, embedding, norm or output head, and counts parameters exactly, including checkpoints stored as FP8, NVFP4, GPTQ, AWQ, MXFP4, MLX or bitsandbytes.
It works across model families: decoder-only LLMs such as Llama, Qwen, DeepSeek, Gemma and gpt-oss, speech models such as Whisper and Parakeet, vision-language models, diffusion transformers and embedding models. Every value on the page says whether it was measured from the weights, read from config.json, or estimated.
How accurate is it? On 510 architectures from the transformers library, 96.8% of tensors get the right role, against 82.3% for the best rule set that reads names only, and 95.0% when every name is replaced by a random token. On 191 popular Hugging Face checkpoints, parameter totals match the Hub's own count for 163 of 170. The method and full results are in the paper.
FAQ
Do I need to download the model?
No. For a Hugging Face repo the page requests only the header bytes of each file, usually under 200 KB in total, even for checkpoints of hundreds of gigabytes.
Does my file leave my computer?
No. Files you drop are parsed in your browser; there is no server. Models you open are kept in your browser, so the next visit does not need to load them again.
How is it different from Netron or Google's Model Explorer?
Those tools draw the operator graph of a model file you already have. This page starts from a repo name, reads only headers, and groups tensors into architectural parts with sizes and precision, so it works on checkpoints with hundreds of billions of parameters without having the files.
Is it free, and can I reuse the code?
Yes. The code is MIT licensed on GitHub. If you use it in research, please cite the paper.
Who made it?
Sattaya Singkul, for the workshop “Squeeze Until It Breaks” at AI Expo Thailand 2026.