LLM Architecture Gallery

---
name: llm-architecture-gallery
description: Source data and metadata management for the LLM Architecture Gallery, a visual catalog of large language model architectures by Sebastian Raschka.
triggers:
  - add a model to the LLM architecture gallery
  - update models.yml for the architecture gallery
  - how do I contribute a new LLM architecture entry
  - edit the LLM gallery metadata
  - add image and metadata for a new language model
  - update model facts in the architecture gallery
  - how does the models.yml schema work
  - submit a new entry to rasbt llm architecture gallery
---

# LLM Architecture Gallery

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

## What This Project Does

The [LLM Architecture Gallery](https://sebastianraschka.com/llm-architecture-gallery/) is a visual catalog of large language model architectures. This repository (`rasbt/llm-architecture-gallery`) is the **source data layer** — it contains the `models.yml` file that drives every gallery card: model names, architecture images, release dates, fact-sheet fields, and external links.

The live site reads from this repo. Contributing here means your model entry appears on the public gallery.

---

## Repository Structure

llm-architecture-gallery/ ├── models.yml # Primary data file — all model metadata └── images/ └── architectures/ # .webp architecture diagrams └── hero/ # Hero/banner images

---

## The `models.yml` Schema

Each top-level key is the **canonical model name** as it appears on the gallery card.

### Minimal Entry

```yaml
GPT-4o:
  image: "/llm-architecture-gallery/images/architectures/gpt-4o.webp"
  date: '2024-05-13'

Full Entry (all known fields)

DeepSeek R1:
  image: "/llm-architecture-gallery/images/architectures/deepseek-v3-r1-671-billion.webp"
  date: '2025-01-20'
  parameters: "671B"
  architecture: "Mixture of Experts (MoE)"
  context_length: "128K"
  organization: "DeepSeek"
  license: "MIT"
  paper: "https://arxiv.org/abs/2501.12948"
  code: "https://github.com/deepseek-ai/DeepSeek-R1"
  notes: "Reasoning model trained via reinforcement learning."

Field Reference

Field	Type	Required	Description
image	string (path)	Yes	Site-relative path to architecture .webp image
date	string (YYYY-MM-DD)	Yes	Public release / paper date
parameters	string	No	E.g. "7B", "671B", "1.5T"
architecture	string	No	E.g. "Transformer", "MoE", "SSM"
context_length	string	No	E.g. "128K", "1M"
organization	string	No	Releasing organization
license	string	No	SPDX identifier or descriptive string
paper	string (URL)	No	arXiv or official paper link
code	string (URL)	No	GitHub or official code link
notes	string	No	Short freeform description

---

Image URL Convention

Images are stored under /llm-architecture-gallery/images/architectures/ as .webp files.

To get a full public URL, prepend https://sebastianraschka.com:

# Stored path in models.yml:
/llm-architecture-gallery/images/architectures/deepseek-v3-r1-671-billion.webp

# Full public URL:
https://sebastianraschka.com/llm-architecture-gallery/images/architectures/deepseek-v3-r1-671-billion.webp

Naming convention for image files:

• lowercase, hyphen-separated
• include organization or model family when helpful
• suffix with parameter count if known
• always .webp format

Examples:

llama-3-70-billion.webp
mistral-7b.webp
deepseek-v3-r1-671-billion.webp
gemma-2-27b.webp

---

Adding a New Model Entry

Step 1 — Prepare the architecture image

Convert your diagram to .webp and place it at:

images/architectures/<model-slug>.webp

Recommended dimensions: ~1200×900px or similar landscape ratio.

Step 2 — Add entry to models.yml

Open models.yml and insert your entry. Entries are not required to be sorted, but keeping them roughly chronological by date helps reviewers.

Llama 3.1 405B:
  image: "/llm-architecture-gallery/images/architectures/llama-3-1-405-billion.webp"
  date: '2024-07-23'
  parameters: "405B"
  architecture: "Transformer (GQA)"
  context_length: "128K"
  organization: "Meta"
  license: "Llama 3.1 Community License"
  paper: "https://arxiv.org/abs/2407.21783"
  code: "https://github.com/meta-llama/llama3"
  notes: "Flagship open-weights model from Meta's Llama 3.1 family."

Step 3 — Validate YAML syntax

# Python — validate the file parses without errors
python3 -c "
import yaml, sys
with open('models.yml') as f:
    data = yaml.safe_load(f)
print(f'OK: {len(data)} models loaded')
for name, meta in data.items():
    if 'image' not in meta:
        print(f'WARNING: {name!r} missing image field')
    if 'date' not in meta:
        print(f'WARNING: {name!r} missing date field')
"

Step 4 — Open a Pull Request

git checkout -b add-llama-3-1-405b
git add models.yml images/architectures/llama-3-1-405-billion.webp
git commit -m "Add Llama 3.1 405B entry"
git push origin add-llama-3-1-405b
# Then open PR at: https://github.com/rasbt/llm-architecture-gallery/pulls

---

Working with models.yml Programmatically

Python — load and query all entries

import yaml
from pathlib import Path
from datetime import date

with open("models.yml") as f:
    models = yaml.safe_load(f)

# List all models sorted by date
sorted_models = sorted(
    models.items(),
    key=lambda kv: kv[1].get("date", "1900-01-01")
)

for name, meta in sorted_models:
    print(f"{meta.get('date', 'unknown')}  {name}  [{meta.get('parameters', '?')}]")

Python — find all MoE models

import yaml

with open("models.yml") as f:
    models = yaml.safe_load(f)

moe_models = {
    name: meta
    for name, meta in models.items()
    if "moe" in meta.get("architecture", "").lower()
    or "mixture" in meta.get("architecture", "").lower()
}

for name, meta in moe_models.items():
    print(f"{name}: {meta.get('architecture')} — {meta.get('parameters', '?')}")

Python — generate a Markdown index

import yaml

with open("models.yml") as f:
    models = yaml.safe_load(f)

BASE_URL = "https://sebastianraschka.com"

lines = ["# Model Index\n"]
for name, meta in sorted(models.items(), key=lambda kv: kv[1].get("date", "")):
    img_url = BASE_URL + meta["image"]
    paper = meta.get("paper", "")
    date = meta.get("date", "")
    params = meta.get("parameters", "")
    org = meta.get("organization", "")
    line = f"- **{name}** ({date}) {org} {params}"
    if paper:
        line += f" — [paper]({paper})"
    lines.append(line)

print("\n".join(lines))

Python — add a new model entry programmatically

import yaml
from collections import OrderedDict

NEW_MODEL = {
    "Phi-4": {
        "image": "/llm-architecture-gallery/images/architectures/phi-4-14b.webp",
        "date": "2024-12-12",
        "parameters": "14B",
        "architecture": "Transformer",
        "context_length": "16K",
        "organization": "Microsoft",
        "license": "MIT",
        "paper": "https://arxiv.org/abs/2412.08905",
        "code": "https://huggingface.co/microsoft/phi-4",
        "notes": "Small language model emphasizing reasoning and STEM."
    }
}

with open("models.yml") as f:
    models = yaml.safe_load(f)

models.update(NEW_MODEL)

with open("models.yml", "w") as f:
    yaml.dump(models, f, allow_unicode=True, sort_keys=False, default_flow_style=False)

print(f"Written. Total models: {len(models)}")

---

Common Patterns

Check for missing required fields across all entries

import yaml

REQUIRED = {"image", "date"}

with open("models.yml") as f:
    models = yaml.safe_load(f)

issues = []
for name, meta in models.items():
    for field in REQUIRED:
        if field not in meta:
            issues.append(f"  {name!r}: missing '{field}'")

if issues:
    print("Validation errors:")
    print("\n".join(issues))
else:
    print(f"All {len(models)} entries valid.")

List all unique organizations

import yaml

with open("models.yml") as f:
    models = yaml.safe_load(f)

orgs = sorted({m.get("organization", "Unknown") for m in models.values()})
for org in orgs:
    print(org)

Count models by architecture family

import yaml
from collections import Counter

with open("models.yml") as f:
    models = yaml.safe_load(f)

arch_counts = Counter(
    m.get("architecture", "Unknown").split("(")[0].strip()
    for m in models.values()
)

for arch, count in arch_counts.most_common():
    print(f"{count:3d}  {arch}")

---

Troubleshooting

YAML parse errors

# Quick check with Python
python3 -c "import yaml; yaml.safe_load(open('models.yml'))"

# Or install yamllint for detailed feedback
pip install yamllint
yamllint models.yml

Common causes:

• Unquoted strings containing : — wrap in quotes: "Transformer: GQA"
• Inconsistent indentation (mix of tabs and spaces)
• Missing quotes around dates that look like floats: use '2024-01-01' not 2024-01-01

Image path not resolving on the live site

Ensure the path starts with /llm-architecture-gallery/images/architectures/ (not a bare filename or full URL). The site prepends https://sebastianraschka.com automatically.

# ✅ Correct
image: "/llm-architecture-gallery/images/architectures/my-model.webp"

# ❌ Wrong — bare filename
image: "my-model.webp"

# ❌ Wrong — full URL
image: "https://sebastianraschka.com/llm-architecture-gallery/images/architectures/my-model.webp"

Date field format

Always quote dates to avoid YAML parsing them as datetime objects:

# ✅ Correct — quoted string
date: '2025-01-20'

# ⚠️ Risky — may parse as datetime depending on YAML loader
date: 2025-01-20

---

Quick Reference

Task	Action
Add new model	Edit models.yml, add .webp to images/architectures/
View live gallery	https://sebastianraschka.com/llm-architecture-gallery/
Full image URL	https://sebastianraschka.com + image field value
Validate YAML	python3 -c "import yaml; yaml.safe_load(open('models.yml'))"
Primary data file	models.yml
License	Apache-2.0