HF Viewer: Visualize Any Hugging Face Model Graph Instantly

HF Viewer is a new web tool that lets you visualize the architecture of any Hugging Face model directly in the browser. No installation, no export step, no config hunting. Just paste a model URL or repo name — e.g., gpt2 — and get an interactive graph showing the high-level structure, from encoder-decoder transformers to sparse MoE reasoning models.
Key Features
- Direct URL magic: Replace
huggingface.cowithhfviewer.comin any model URL to view it instantly. - Granularity levels: Zoom from overview down to specific sub-structures, such as attention blocks, vision encoders, or expert routing.
- Model family comparison: Compare related models side-by-side with synchronized pan/zoom — currently showcased for the Gemma 4 family.
- Embed in model cards: Press the Embed button to get an iframe snippet for your own model card.
How to Use
Navigate to hfviewer.com, paste a Hugging Face model URL or repo name in the input box, and click "Visualize Model". Alternatively, manually replace huggingface.co with hfviewer.com in the URL bar.
For example, to visualize GPT-2: open https://hfviewer.com/gpt2.
Use Cases
The tool is designed for developers and ML engineers who need to quickly understand a model's architecture without reading through config files or source code. It supports a range of popular models including:
- Qwen/Qwen3.5-0.8B — small instruction-tuned LLM
- google/vit-base-patch16-224 — vision backbone
- openai/clip-vit-base-patch32 — dual encoder
- t5-small — encoder-decoder
- nvidia/parakeet-tdt-0.6b-v3 — streaming Conformer-TDT speech recognizer
Interactive Blog Format
On the Gemma 4 family page, the blog text and graph are linked. You can read about an architectural decision and jump into the corresponding part of the graph, then return to the article with surrounding context intact. This graph-to-text loop offers a new way to communicate ML architecture.
HF Viewer is released as a free community tool by the Embedl team.
📖 Read the full source: HN AI Agents
👀 See Also

pop-pay MCP server adds payment guardrails for Claude Code agents
pop-pay is an MCP server that lets Claude Code agents handle purchases without exposing credit card numbers. It uses CDP injection to place virtual card credentials directly into payment iframes, with Claude only receiving masked confirmation numbers.

Claude Code Hook Monitors WIP Accumulation in AI Coding Workflows
A developer built a UserPromptSubmit hook for Claude Code that surfaces work-in-progress accumulation across four queues: uncommitted changes over 200 lines, three or more unpushed commits, pushed commits without changeset files, and release PRs open longer than 24 hours.

KubeShark: A Kubernetes Skill for Claude Code and Codex to Catch Hallucinated YAML
KubeShark is a failure-mode-first Kubernetes skill for Claude Code and Codex that catches deprecated APIs, misconfigured probes, broken selectors, and other common AI-generated mistakes before they hit production.

Deploy OpenClaw on VPS with a One-Command CLI
A Reddit user claims to have developed a CLI that deploys OpenClaw on a $4.99/month VPS with a single command, offering a cost-effective alternative to using Mac Minis.