Modelmap: Interactive Visualizations of Hugging Face Model Architectures

Modelmap is a tool designed to provide an interactive, animated visualization of the network architecture of any model hosted on Hugging Face. By entering a Hugging Face model ID, users can generate a living map of the model's network without needing to download the actual model weights.

Technical Implementation of Architecture Mapping

Modelmap generates its architecture graphs using a combination of meta-device instantiation and traced fake forward passes. This approach allows the tool to determine the structural layout and tensor shapes of a model without the actual weights, making the visualization process nearly instant and bandwidth-efficient.

  • Meta-device Instantiation: The tool instantiates the model structure on a meta-device, which allows it to define the architecture without allocating memory for the actual parameters.
  • Traced Fake Forward Pass: The tool performs a trace of a simulated forward pass to determine the exact shapes of the tensors as they flow through the network.

Supported Model Types and Capabilities

Modelmap supports a wide range of architectures, including classic reference architectures and modern large language models (LLMs). Users can also compare two different models side-by-side using the same interface.

Reference Architectures

Modelmap provides instant access to visualizations for several foundational models, such as:

  • GPT-2 (124M): A classic decoder-only architecture.
  • BERT (base-uncased): The original bidirectional encoder-only transformer.
  • T5/Qwen3-8B: Modern dense LLMs featuring Grouped-Query Attention (GQA), RMSNorm, and gated MLP layers.
  • DeepSeek-V3.1 (671B): A massive Mixture-of-Experts (MoE) model with multi-head latent attention.
  • Qwen3-235B-A22B: An MoE model with 128 experts per layer.

Advanced Model Categories

The tool also supports vision-language models, such as Qwen2.5-VL-3B-Instruct, where a vision tower feeds into an LLM, and various image-text-to-text models like Qwen3.8-27B and Muse-Glimmer-30B.

User Experience and Access Control

Modelmap provides a streamlined experience for debugging and understanding model structures. Users can actually use the tool to estimate the cost of serving a model based on its architecture.

  • Public Repositories: Any public Hugging Face repository can be visualized immediately.
  • Gated Repositories: For gated models, users can provide a Hugging Face token to gain access to the architecture map.
  • Comparison Tool: The tool includes a comparison feature (triggered by ⌘K) that allows users to visualize the architecture of two models (e.g., Qwen2.5-7B vs Qwen3-8B) for direct structural comparison.

Community Feedback

The community has noted the tool's utility for debugging fine-tunes and LoRAs, as some users have previously relied on LLMs like Claude to "walk model architectures" to debugging purposes.

"I’ve had Claude walk model architectures to debug Loras and fine tunes, but this is delightful"

Other users have pointed to similar tools like hfviewer.com, noting differences in design aesthetics.

Sources

Related

  • Dispatch
    Introducing HF viewer: An Interactive Visualizer for Hugging Face Models

    HF viewer is a new interactive tool designed to visualize any Hugging Face model by simply pasting its URL, offering detailed insights into model architecture at multiple granularities. This tool aims to enhance understanding and exploration of complex machine learning models for developers and researchers.

  • Dispatch
    Hugging Face Transformers: Standardizing Model Definitions for Ecosystem Interoperability

    Hugging Face is positioning the Transformers library as the central pivot for model definitions to ensure that any architecture supported by Transformers is automatically compatible with the broader ML ecosystem, including inference engines and training frameworks.

  • Dispatch
    Hugging Face Spaces and Gradio Integration

    Hugging Face has integrated Gradio into Spaces, allowing users to easily host and showcase machine learning model demos using the Inference API or custom model checkpoints.

  • Dispatch
    Hugging Face and Renumics Spotlight Integration for Scalable Data Inspection

    Hugging Face has integrated with Renumics Spotlight to enable interactive, one-line-of-code visualization and inspection of ML datasets, including support for multimodal data and model results.

  • Project
    OpenEnvision/Awesome-Multimodal-Modeling

    Awesome‑Multimodal‑Modeling is a community‑curated “awesome‑list” that surveys multimodal AI models. It defines four clear families—Traditional, MLLMs, Unified, and Native—explains their architectural differences, and links to open‑source model collections on Hugging Face. The repo serves researchers and engineers who need a structured view of the rapidly evolving multimodal landscape.