# AI Model Explorer

> Free, open-source (MIT) web tool by Sattaya Singkul that splits any AI model architecture into its parts and shows the real weights, compute, TFLite op count and precision of each part. Thai and English. https://joessattes.github.io/ai-model-explorer/

AI Model Explorer reads a model from four kinds of input and works out its anatomy automatically:

- a Hugging Face repository id: only the file list, the safetensors headers (by HTTP range request), shard indexes and config.json are read; weights are never downloaded and no model code is run;
- a `.safetensors` file: its header;
- a `.tflite` (LiteRT) file: the real op graph is parsed in the browser (ops, tensor shapes, dtypes, quantize/dequantize "seams");
- a pasted `config.json`: the structure is synthesized for common families.

From tensor names, shapes and dtypes it recovers components (vision tower, projector, LLM backbone, transducer prediction and joint networks, diffusion denoisers), repeated layer stacks and block types, what each part does (attention including GQA, latent and cross-attention; dense and gated feed-forward; mixture of experts; SSM / linear attention; conformer convolution; adaLN modulation; recurrent cells; multi-token-prediction layers), how the parts connect, and exact parameter counts under NVFP4, FP8, GPTQ, AWQ, MXFP4, MLX and bitsandbytes storage. Every value is tagged as measured from the weights, read from config and checked against the weights, or estimated.

Evaluation: 96.8% per-tensor role accuracy on 510 transformers architectures (95.0% with names hidden; name-reading baselines reach 82.3%, and llama.cpp's tensor-name table covers about a fifth of them), exact parameter totals for all of them; on 191 popular Hugging Face checkpoints, totals agree with the Hub's own count for 163 of 170. Paper: https://www.researchgate.net/publication/417293353_Reading_the_Architecture_off_the_Weights_Manifest-Only_Recovery_of_Neural_Network_Anatomy · benchmark page: https://joessattes.github.io/ai-model-explorer/bench.html

## What it shows
- Anatomy view: components in data-path order, a strip per stack coloured by layer type, the parameter share of each region, the evidence behind each label, and a panel of values with their provenance.
- Big picture: the model as numbered parts (input → embedding → encoder/decoder stacks → output head), with a play-through that shows each part's input and output, a zoom into one layer, and the decoder's per-token loop.
- Lenses: weights (bytes, params), compute (FLOPs), TFLite op count, precision (fp32/fp16/bf16/int8/int4), fragility (output error when one part is quantized to int4, where measured).
- Layer map drawn to scale, including weight bytes read per input (the decoder re-reads its weights every token).
- Concepts with numbers from the selected model (attention, feed-forward, KV cache, tied embedding, seams) and a quantization playground.
- Report with insights and a Markdown table to copy.

## Templates included
- biodatlab/distill-whisper-th-medium (Thai speech recognition) — real .tflite graph of the architecture, 1,616 ops
- Qwen/Qwen3-ASR-1.7B (audio encoder → LLM) and Qwen/Qwen3.8-27B (vision encoder → hybrid linear/full-attention LLM, 64 layers, MTP head) — structure built from the official config.json
- meta-llama/Llama-3.2-1B, openai-community/gpt2, google-bert/bert-base-uncased, google/vit-base-patch16-224 — structure from the transformers implementation
- Whisper rehearsal models in fp32, weight-only int8, static int8 and dynamic int4 for precision comparisons
- NVIDIA Nemotron 3.5 Super VL, Transcribe, ASR streaming and Embed NVFP4; Parakeet TDT 0.6B v3; FastH3 Omni FP8; Kumo Tabular and Relational, analysed automatically

## Author
Sattaya Singkul (Jo) — Researcher, speech and edge AI. Portfolio: https://joessattes.github.io/Sattaya-AI-Website/ · GitHub: https://github.com/JoesSattes
Built for the workshop "Squeeze Until It Breaks — A Hands-On Edge Inference Contest", AI Expo Thailand 2026.

## Links
- App: https://joessattes.github.io/ai-model-explorer/ (Thai) · https://joessattes.github.io/ai-model-explorer/en/ (English)
- Python package: https://pypi.org/project/ai-model-explorer/ (`pip install ai-model-explorer`; CLI `ai-model-explorer inspect|insights|memory|compare|plot|report|open`; API `from ai_model_explorer import explore, compare`: active parameters, parameter breakdown, KV-cache and memory estimates, diagrams, HTML reports)
- Usage guide (web, CLI, Python, JSON fields): https://github.com/JoesSattes/ai-model-explorer/blob/main/docs/usage.md
- How it works: https://github.com/JoesSattes/ai-model-explorer/blob/main/docs/how-it-works.md
- Source: https://github.com/JoesSattes/ai-model-explorer
- Paper: https://www.researchgate.net/publication/417293353_Reading_the_Architecture_off_the_Weights_Manifest-Only_Recovery_of_Neural_Network_Anatomy (PDF: https://joessattes.github.io/ai-model-explorer/paper/manifest-anatomy-singkul-2026.pdf)
- Full text for AI tools: https://joessattes.github.io/ai-model-explorer/llms-full.txt
- Hub benchmark: https://joessattes.github.io/ai-model-explorer/bench.html
- License: MIT, Copyright (c) 2026 Sattaya Singkul

Cite: Sattaya Singkul. 2026. Reading the Architecture off the Weights: Manifest-Only Recovery of Neural Network Anatomy. Preprint, ResearchGate. https://www.researchgate.net/publication/417293353_Reading_the_Architecture_off_the_Weights_Manifest-Only_Recovery_of_Neural_Network_Anatomy