A newer version of the Gradio SDK is available: 6.12.0
title: Neural Model Analyzer
emoji: 🔬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.9.0
app_file: app.py
pinned: true
license: mit
short_description: 'Advanced model introspection: 150+ archs, weights'
tags:
- model-analysis
- pytorch
- visualization
- deep-learning
\U0001f52c Neural Model Analyzer
Upload any model file and get a deep analysis — architecture detection (150+ families), weight distributions, layer connectivity, memory profiling, and more. Runs on free HuggingFace CPU Spaces.
Supported Formats
| Format | Extension | Notes |
|---|---|---|
| PyTorch | .pth, .pt |
State dicts, full models, checkpoints |
| HuggingFace | .bin |
Standard HF model format |
| SafeTensors | .safetensors |
Fast, safe tensor format |
| ONNX | .onnx |
With input/output specs |
150+ Detected Architecture Families
NLP Encoders: BERT, RoBERTa, DistilBERT, ALBERT, ELECTRA, DeBERTa, XLM-R, Longformer, BigBird, MPNet, ConvBERT, MobileBERT, ...
Sentence Embeddings: Sentence-BERT, E5, BGE, GTE, Instructor, Jina, ...
Causal LM / Decoders: GPT-2, GPT-NeoX, GPT-J, LLaMA 1/2/3, Mistral, Mixtral (MoE), Phi, Qwen, Falcon, BLOOM, OPT, Mamba, RWKV, ChatGLM, DeepSeek, Yi, Baichuan, InternLM, StableLM, DBRX, OLMo, CodeLlama, StarCoder, ...
Seq2Seq: T5, Flan-T5, BART, mBART, Pegasus, MarianMT, NLLB, CodeT5, LED, ...
Vision: ViT, DeiT, BEiT, Swin, ConvNeXt, ResNet, EfficientNet, MobileNet, DINOv2, EVA, PVT, MaxViT, MLP-Mixer, ...
Detection & Segmentation: DETR, RT-DETR, YOLO, Faster R-CNN, SAM, SAM 2, SegFormer, Mask2Former, DeepLab, ...
Audio: Whisper, Wav2Vec2, HuBERT, WavLM, SpeechT5, Bark, VITS, MusicGen, AudioLDM, CLAP, ...
Multimodal: CLIP, OpenCLIP, SigLIP, BLIP, BLIP-2, LLaVA, InternVL, PaliGemma, CogVLM, Florence, Kosmos, ...
Generative: Stable Diffusion, SDXL, SD3, FLUX, DiT, PixArt, ControlNet, IP-Adapter, VQGAN, StyleGAN, ESRGAN, ...
Science: ESM (protein), AlphaFold-style, GNN, SchNet, Graphormer, MolBART, ...
PEFT: LoRA, QLoRA, IA3, Prefix Tuning, Prompt Tuning, Adapters, ...
And more: Time series (PatchTST, Informer), Video (VideoMAE, CogVideo), Document (Donut, TrOCR), RL (Decision Transformer), Quantized (GPTQ, AWQ, BnB), ...
Analysis Tabs
- Summary & Architecture — File info, param count, auto-detected architecture with confidence, inferred config (hidden size, layers, heads, vocab, etc.)
- Layer Tree — Full hierarchical view with shapes and sizes
- Types & Connections — Layer type inference + connectivity (direct, reshape, skip/residual)
- Weight Statistics — Per-tensor mean, std, sparsity, health warnings
- Distributions — Histograms per layer
- Module Sizes — Bar chart by module
- Heatmap — Normalized statistics across layers
- Memory — Dtype breakdown + largest tensors
- Depth Profile — Parameter count across network depth